US20130170564A1 - Encoding of a picture in a video sequence by example-based data pruning using intra-frame patch similarity - Google Patents

Abstract

Method and apparatus for encoding a picture in video sequence are disclosed. An apparatus includes a library creator for creating a first library from an original version of the picture and a second library from a reconstructed version of the picture. Each library includes high resolution replacement patches for replacing pruned blocks during a recovery of a pruned version of the picture. A pruner generates the pruned version from the first library. A metadata generator generates metadata from the second library for recovering the pruned version. An encoder encodes the pruned version and metadata. The first library includes patch clusters. The pruned version is generated by dividing the original version into overlapping blocks and searching for candidate patch clusters for each block. A patch dependency graph having nodes and edges is used for the searching. Each node represents a respective block, and each edge represents a respective dependency of the respective block.

Description

• This application claims the benefit of U.S. Provisional Application Ser. No. 61/403,107 entitled EXAMPLE-BASED DATA PRUNING USING INTRA-FRAME PATCH SIMILARITY filed on Sep. 10, 2010 (Technicolor Docket No. PU100196).
• This application is related to the following co-pending, commonly-owned, patent applications:
• • (1) International (PCT) Patent Application Serial No. PCT/US11/000107 entitled A SAMPLING-BASED SUPER-RESOLUTION APPROACH FOR EFFICIENT VIDEO COMPRESSION filed on Jan. 20, 2011 (Technicolor Docket No. PU100004);
  • (2) International (PCT) Patent Application Serial No. PCT/US11/000117 entitled DATA PRUNING FOR VIDEO COMPRESSION USING EXAMPLE-BASED SUPER-RESOLUTION filed on Jan. 21, 2011 (Technicolor Docket No. PU100014);
  • (3) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR ENCODING VIDEO SIGNALS USING MOTION COMPENSATED EXAMPLE-BASED SUPER-RESOLUTION FOR VIDEO COMPRESSION filed on Sep. XX, 2011 (Technicolor Docket No. PU100190);
  • (4) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR DECODING VIDEO SIGNALS USING MOTION COMPENSATED EXAMPLE-BASED SUPER-RESOLUTION FOR VIDEO COMPRESSION filed on Sep. XX, 2011 (Technicolor Docket No. PU100266);
  • (5) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR ENCODING VIDEO SIGNALS USING EXAMPLE-BASED DATA PRUNING FOR IMPROVED VIDEO COMPRESSION EFFICIENCY filed on Sep. XX, 2011 (Technicolor Docket No. PU100193);
  • (6) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR DECODING VIDEO SIGNALS USING EXAMPLE-BASED DATA PRUNING FOR IMPROVED VIDEO COMPRESSION EFFICIENCY filed on Sep. XX, 2011 (Technicolor Docket No. PU100267);
  • (7) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR ENCODING VIDEO SIGNALS FOR BLOCK-BASED MIXED-RESOLUTION DATA PRUNING filed on Sep. XX, 2011 (Technicolor Docket No. PU100194);
  • (8) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR DECODING VIDEO SIGNALS FOR BLOCK-BASED MIXED-RESOLUTION DATA PRUNING filed on Sep. XX, 2011 (Technicolor Docket No. PU100268);
  • (9) International (PCT) Patent Application Serial No. XXXX entitled METHODS AND APPARATUS FOR EFFICIENT REFERENCE DATA ENCODING FOR VIDEO COMPRESSION BY IMAGE CONTENT BASED SEARCH AND RANKING filed on Sep. XX, 2011 (Technicolor Docket No. PU100195);
  • (10) International (PCT) Patent Application Serial No. XXXX entitled METHOD AND APPARATUS FOR EFFICIENT REFERENCE DATA DECODING FOR VIDEO COMPRESSION BY IMAGE CONTENT BASED SEARCH AND RANKING filed on Sep. XX, 2011 (Technicolor Docket No. PU110106);
  • (11) International (PCT) Patent Application Serial No. XXXX entitled METHOD AND APPARATUS FOR DECODING VIDEO SIGNALS WITH EXAMPLE-BASED DATA PRUNING USING INTRA-FRAME PATCH SIMILARITY filed on Sep. XX, 2011 (Technicolor Docket No. PU100269);
  • (12) International (PCT) Patent Application Serial No. XXXX entitled PRUNING DECISION OPTIMIZATION IN EXAMPLE-BASED DATA PRUNING COMPRESSION filed on Sep. XX, 2011 (Technicolor Docket No. PU10197).
• The present principles relate generally to video encoding and decoding and, more particularly, to methods and apparatus for example-based data pruning using intra-frame patch similarity.
• Data pruning is a video preprocessing technology that achieves better video coding efficiency by removing part of the input video data before the input video data is encoded. The removed video data is recovered at the decoder side by inferring the removed video data from the decoded data. One example of data pruning is image line removal, which removes some of the horizontal and vertical scan lines in the input video.
• In a first approach, a new data pruning method called example-based data pruning is employed, in which external videos or video frames that have been previously transmitted to the decoder side are used to train an example patch library. The patch library is then used to prune and recover the video data.
• There have been several efforts to explore using data pruning to increase compression efficiency. For example, in a second approach and a third approach, a texture replacement based method is used to remove texture regions at the encoder side, and re-synthesize the texture regions at the decoder side. Compression efficiency is gained because only synthesis parameters are sent to the decoder, which are smaller than the regular transformation coefficients. In a fourth approach and a fifth approach, spatio-temporal texture synthesis and edge-based inpainting are used to remove some of the regions at the encoder side, and the removed content is recovered at the decoder side, with the help of metadata such as region masks. However, the fourth and fifth approaches need to modify the encoder and decoder so that the encoder/decoder can selectively perform encoding/decoding for some of the regions using the region masks. Therefore, it is not exactly an out-of-loop approach (i.e., the encoder and decoder need to be modified). In a sixth approach, a line removal based method is proposed to rescale a video to a smaller size by selectively removing some of the horizontal or vertical lines in the video with a least-square minimization framework. The sixth approach is an out-of-loop approach, and does not require modification of the encoder/decoder. However, completely removing certain horizontal and vertical lines may result in loss of information or details for some videos.
• Some preliminary research on data pruning for video compression has been conducted. For example, in a seventh approach, a data pruning scheme using sampling-based super-resolution is presented. The full resolution frame is sampled into several smaller-sized frames, therefore reducing the spatial size of the original video. At the decoder side, the high-resolution frame is re-synthesized from the downsampled frames with the help of metadata received from the encoder side. In an eighth approach, an example-based super-resolution based method for data pruning is presented. A representative patch library is trained from the original video. Afterwards, the video is downsized to a smaller size. The downsized video and the patch library are sent to the decoder side. The recovery process at the decoder side super-resolves the downsized video by example-based super-resolution using the patch library. However because there is substantial redundancy between the patch library and downsized frames, it has been discovered that it may be difficult to achieve compression gain using the eighth approach.
• In the aforementioned first approach, an example-based data pruning method creates a patch library using the video frames that have been sent to the decoder side and uses the patch library to prune and recover video frames. However, this method does not consider the intra-frame patch dependency, which may happen if there are repetitive textures or patterns in a video frame.
• In the International Organization for Standardization/International Electrotechnical Commission (ISO/IEC) Moving Picture Experts Group-4 (MPEG-4)Part 10 Advanced Video Coding (AVC) Standard/International Telecommunication Union, Telecommunication Sector (ITU-T) H.264 Recommendation (hereinafter the “MPEG-4 AVC Standard”), intra-frame block prediction is realized by block prediction from the neighboring blocks. However, long-range similarity of non-neighboring blocks is not exploited to increase compression efficiency.
• These and other drawbacks and disadvantages of these approaches are addressed by the present principles, which are directed to methods and apparatus for example-based data pruning using intra-frame patch similarity.
• According to an aspect of the present principles, there is provided an apparatus for encoding a picture in a video sequence. The apparatus includes a patch library creator for creating a first patch library from an original version of the picture and a second patch library from a reconstructed version of the picture. Each of the first patch library and the second patch library includes a plurality of high resolution replacement patches for replacing one or more pruned blocks during a recovery of a pruned version of the picture. The apparatus also includes a pruner for generating the pruned version of the picture from the first patch library. The apparatus further includes a metadata generator for generating metadata from the second patch library. The metadata is for recovering the pruned version of the picture. The apparatus additionally includes an encoder for encoding the pruned version of the picture and the metadata. The first patch library includes a plurality of patch clusters, and the pruned version of the picture is generated by dividing the original version of the picture into a plurality of overlapping blocks, searching for candidate patch clusters from among the plurality of patch clusters for each of the plurality of overlapping blocks based on respective distance metrics from each of the plurality of overlapping blocks to respective centers of each of the plurality of patch clusters, identifying a best matching patch from the candidate patch clusters based on one or more criterion, and pruning a corresponding one of the plurality of overlapping blocks to obtain a pruned block there for when a difference between the corresponding one of the plurality of overlapping blocks and the best matching patch is less than a threshold difference. A patch dependency graph having a plurality of nodes and a plurality of edges is used for the searching. Each of the plurality of nodes represents a respective one of the plurality of overlapping blocks, and each of the plurality of edges represents a respective dependency of at least the respective one of the plurality of overlapping blocks.
• According to another aspect of the present principles, there is provided a method for encoding a picture in a video sequence. The method includes creating a first patch library from an original version of the picture and a second patch library from a reconstructed version of the picture. Each of the first patch library and the second patch library includes a plurality of high resolution replacement patches for replacing one or more pruned blocks during a recovery of a pruned version of the picture. The method also includes generating the pruned version of the picture from the first patch library. The method further includes generating metadata from the second patch library. The metadata is for recovering the pruned version of the picture. The method additionally includes encoding the pruned version of the picture and the metadata. The first patch library includes a plurality of patch clusters, and the pruned version of the picture is generated by dividing the original version of the picture into a plurality of overlapping blocks, searching for candidate patch clusters from among the plurality of patch clusters for each of the plurality of overlapping blocks based on respective distance metrics from each of the plurality of overlapping blocks to respective centers of each of the plurality of patch clusters, identifying a best matching patch from the candidate patch clusters based on one or more criterion, and pruning a corresponding one of the plurality of overlapping blocks to obtain a pruned block there for when a difference between the corresponding one of the plurality of overlapping blocks and the best matching patch is less than a threshold difference. A patch dependency graph having a plurality of nodes and a plurality of edges is used for the searching. Each of the plurality of nodes represents a respective one of the plurality of overlapping blocks, and each of the plurality of edges represents a respective dependency of at least the respective one of the plurality of overlapping blocks.
• According to still another aspect of the present principles, there is provided an apparatus for recovering a pruned version of a picture in a video sequence. The apparatus includes a divider for dividing the pruned version of the picture into a plurality of non-overlapping blocks. The apparatus also includes a metadata decoder for decoding metadata for use in recovering the pruned version of the picture. The apparatus further includes a patch library creator for creating a patch library from a reconstructed version of the picture. The patch library includes a plurality of high resolution replacement patches for replacing the one or more pruned blocks during a recovery of the pruned version of the picture. The apparatus additionally includes a search and replacement device for performing a searching process using the metadata to find a corresponding patch for a respective one of the one or more pruned blocks from among the plurality of non-overlapping blocks and replace the respective one of the one or more pruned blocks with the corresponding patch. The signature is respectively created for each of the one or more pruned blocks, and the pruned version of the picture is recovered by comparing respective distance metrics from signatures for each of the plurality of high resolution patches to signatures for each of the one or more pruned blocks, sorting the respective distance metrics to obtain a rank list for each of the one or more pruned blocks, wherein a rank number in the rank list for a particular one of the one or more pruned blocks is used to retrieve a corresponding one of the plurality of high resolution patches in the patch library to be used to replace the particular one of the one or more pruned blocks. A patch dependency graph having a plurality of nodes and a plurality of edges is used to recover the pruned version of the picture. Each of the plurality of nodes represents a respective one of the plurality of overlapping blocks, and each of the plurality of edges represents a respective dependency of at least the respective one of the plurality of overlapping blocks.
• According to a further aspect of the present principles, there is provided a method for recovering a pruned version of a picture in a video sequence. The method includes dividing the pruned version of the picture into a plurality of non-overlapping blocks. The method also includes decoding metadata for use in recovering the pruned version of the picture. The method further includes creating a patch library from a reconstructed version of the picture. The patch library includes a plurality of high resolution replacement patches for replacing the one or more pruned blocks during a recovery of the pruned version of the picture. The method additionally includes performing a searching process using the metadata to find a corresponding patch for a respective one of the one or more pruned blocks from among the plurality of non-overlapping blocks and replace the respective one of the one or more pruned blocks with the corresponding patch. The signature is respectively created for each of the one or more pruned blocks, and the pruned version of the picture is recovered by comparing respective distance metrics from signatures for each of the plurality of high resolution patches to signatures for each of the one or more pruned blocks, sorting the respective distance metrics to obtain a rank list for each of the one or more pruned blocks, wherein a rank number in the rank list for a particular one of the one or more pruned blocks is used to retrieve a corresponding one of the plurality of high resolution patches in the patch library to be used to replace the particular one of the one or more pruned blocks. A patch dependency graph having a plurality of nodes and a plurality of edges is used to recover the pruned version of the picture. Each of the plurality of nodes represents a respective one of the plurality of overlapping blocks, and each of the plurality of edges represents a respective dependency of at least the respective one of the plurality of overlapping blocks.
• According to a still further aspect of the present principles, there is provided an apparatus for encoding a picture in a video sequence. The apparatus includes means for creating a first patch library from an original version of the picture and a second patch library from a reconstructed version of the picture. Each of the first patch library and the second patch library includes a plurality of high resolution replacement patches for replacing one or more pruned blocks during a recovery of a pruned version of the picture. The apparatus also includes means for generating the pruned version of the picture from the first patch library. The apparatus further includes means for generating metadata from the second patch library, the metadata for recovering the pruned version of the picture. The apparatus additionally includes means for encoding the pruned version of the picture and the metadata. The first patch library includes a plurality of patch clusters, and the pruned version of the picture is generated by dividing the original version of the picture into a plurality of overlapping blocks, searching for candidate patch clusters from among the plurality of patch clusters for each of the plurality of overlapping blocks based on respective distance metrics from each of the plurality of overlapping blocks to respective centers of each of the plurality of patch clusters, identifying a best matching patch from the candidate patch clusters based on one or more criterion, and pruning a corresponding one of the plurality of overlapping blocks to obtain a pruned block there for when a difference between the corresponding one of the plurality of overlapping blocks and the best matching patch is less than a threshold difference. A patch dependency graph having a plurality of nodes and a plurality of edges is used for the searching. Each of the plurality of nodes represents a respective one of the plurality of overlapping blocks, and each of the plurality of edges represents a respective dependency of at least the respective one of the plurality of overlapping blocks.
• According to an additional aspect of the present principles, there is provided an apparatus for recovering a pruned version of a picture in a video sequence. The apparatus includes means for dividing the pruned version of the picture into a plurality of non-overlapping blocks. The apparatus also includes means for decoding metadata for use in recovering the pruned version of the picture. The apparatus further includes means for creating a patch library from a reconstructed version of the picture. The patch library includes a plurality of high resolution replacement patches for replacing the one or more pruned blocks during a recovery of the pruned version of the picture. The apparatus additionally includes means for performing a searching process using the metadata to find a corresponding patch for a respective one of the one or more pruned blocks from among the plurality of non-overlapping blocks and replace the respective one of the one or more pruned blocks with the corresponding patch. The signature is respectively created for each of the one or more pruned blocks, and the pruned version of the picture is recovered by comparing respective distance metrics from signatures for each of the plurality of high resolution patches to signatures for each of the one or more pruned blocks, sorting the respective distance metrics to obtain a rank list for each of the one or more pruned blocks, wherein a rank number in the rank list for a particular one of the one or more pruned blocks is used to retrieve a corresponding one of the plurality of high resolution patches in the patch library to be used to replace the particular one of the one or more pruned blocks. A patch dependency graph having a plurality of nodes and a plurality of edges is used to recover the pruned version of the picture. Each of the plurality of nodes represents a respective one of the plurality of overlapping blocks, and each of the plurality of edges represents a respective dependency of at least the respective one of the plurality of overlapping blocks.
• These and other aspects, features and advantages of the present principles will become apparent from the following detailed description of exemplary embodiments, which is to be read in connection with the accompanying drawings.
• The present principles may be better understood in accordance with the following exemplary figures, in which:
• FIG. 1 is a block diagram showing an exemplary example-based data pruning system using intra-frame patch similarity, in accordance with an embodiment of the present principles;
• FIG. 2 is a block diagram showing an exemplary video encoder to which the present principles may be applied, in accordance with an embodiment of the present principles;
• FIG. 3 is a block diagram showing an exemplary video decoder to which the present principles may be applied, in accordance with an embodiment of the present principles;
• FIG. 4 is a block diagram showing an exemplary first portion for performing encoder side processing in an example-based data pruning system using intra-frame patch similarity, in accordance with an embodiment of the present principles;
• FIG. 5 is a block diagram showing an exemplary method for clustering and patch library creation, in accordance with an embodiment of the present principles;
• FIG. 6 is a block diagram showing an exemplary patch library and corresponding clusters, in accordance with an embodiment of the present principles;
• FIG. 7 is a diagram showing an exemplary signature vector, in accordance with an embodiment of the present principles;
• FIG. 8 is a block diagram showing an exemplary second portion for performing encoder side processing in an example-based data pruning system using intra-frame patch similarity, in accordance with an embodiment of the present principles;
• FIG. 9 is a flow diagram showing an exemplary method for video frame pruning, in accordance with an embodiment of the present principles;
• FIG. 10 is a block diagram showing a patch search process, in accordance with an embodiment of the present principles;
• FIG. 11 is a diagram showing a block search area in a process of pruning for causal recovery, in accordance with an embodiment of the present principles;
• FIG. 12 is a diagram showing an exemplary block dependency graph, in accordance with an embodiment of the present principles;
• FIG. 13 is a flow diagram showing an exemplary method for obtaining a recovery sequence, in accordance with an embodiment of the present principles;
• FIG. 14 is a diagram showing an exemplary evolution of the dependency graph using the pruning algorithm, in accordance with an embodiment of the present principles;
• FIG. 15 is a diagram showing an exemplary mixed-resolution frame, in accordance with an embodiment of the present principles;
• FIG. 16 is a flow diagram showing an exemplary method for encoding metadata, in accordance with an embodiment of the present principles;
• FIG. 17 is a flow diagram showing an example method for encoding pruned block IDs, in accordance with an embodiment of the present principles;
• FIG. 18 is a flow diagram showing an exemplary method for encoding a patch index, in accordance with an embodiment of the present principles;
• FIG. 19 is a flow diagram showing an exemplary method for decoding a patch index, in accordance with an embodiment of the present principles;
• FIG. 20 is a diagram showing an exemplary block ID, in accordance with an embodiment of the present principles;
• FIG. 21 is a flow diagram showing an exemplary method for pruning sequent frames, in accordance with an embodiment of the present principles;
• FIG. 22 is a diagram showing an exemplary motion vector for a pruned block, in accordance with an embodiment of the present principles;
• FIG. 23 is a flow diagram showing an exemplary method for decoding metadata, in accordance with an embodiment of the present principles;
• FIG. 24 is a flow diagram showing an exemplary method for decoding pruned block IDs, in accordance with an embodiment of the present principles;
• FIG. 25 is a block diagram showing an exemplary apparatus for performing decoder side processing for example-based data pruning using intra-frame patch similarity, in accordance with an embodiment of the present principles;
• FIG. 26 is a flow diagram showing an exemplary method for recovering a pruned frame, in accordance with an embodiment of the present principles; and
• FIG. 27 is a flow diagram showing an exemplary method for recovering subsequent frames, in accordance with an embodiment of the present principles.
• The present principles are directed to methods and apparatus for example-based data pruning using intra-frame patch similarity.
• The present description illustrates the present principles. It will thus be appreciated that those skilled in the art will be able to devise various arrangements that, although not explicitly described or shown herein, embody the present principles and are included within its spirit and scope.
• All examples and conditional language recited herein are intended for pedagogical purposes to aid the reader in understanding the present principles and the concepts contributed by the inventor(s) to furthering the art, and are to be construed as being without limitation to such specifically recited examples and conditions.
• Moreover, all statements herein reciting principles, aspects, and embodiments of the present principles, as well as specific examples thereof, are intended to encompass both structural and functional equivalents thereof. Additionally, it is intended that such equivalents include both currently known equivalents as well as equivalents developed in the future, i.e., any elements developed that perform the same function, regardless of structure.
• Thus, for example, it will be appreciated by those skilled in the art that the block diagrams presented herein represent conceptual views of illustrative circuitry embodying the present principles. Similarly, it will be appreciated that any flow charts, flow diagrams, state transition diagrams, pseudocode, and the like represent various processes which may be substantially represented in computer readable media and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.
• The functions of the various elements shown in the figures may be provided through the use of dedicated hardware as well as hardware capable of executing software in association with appropriate software. When provided by a processor, the functions may be provided by a single dedicated processor, by a single shared processor, or by a plurality of individual processors, some of which may be shared. Moreover, explicit use of the term “processor” or “controller” should not be construed to refer exclusively to hardware capable of executing software, and may implicitly include, without limitation, digital signal processor (“DSP”) hardware, read-only memory (“ROM”) for storing software, random access memory (“RAM”), and non-volatile storage.
• Other hardware, conventional and/or custom, may also be included. Similarly, any switches shown in the figures are conceptual only. Their function may be carried out through the operation of program logic, through dedicated logic, through the interaction of program control and dedicated logic, or even manually, the particular technique being selectable by the implementer as more specifically understood from the context.
• In the claims hereof, any element expressed as a means for performing a specified function is intended to encompass any way of performing that function including, for example, a) a combination of circuit elements that performs that function or b) software in any form, including, therefore, firmware, microcode or the like, combined with appropriate circuitry for executing that software to perform the function. The present principles as defined by such claims reside in the fact that the functionalities provided by the various recited means are combined and brought together in the manner which the claims call for. It is thus regarded that any means that can provide those functionalities are equivalent to those shown herein.
• Reference in the specification to “one embodiment” or “an embodiment” of the present principles, as well as other variations thereof, means that a particular feature, structure, characteristic, and so forth described in connection with the embodiment is included in at least one embodiment of the present principles. Thus, the appearances of the phrase “in one embodiment” or “in an embodiment”, as well any other variations, appearing in various places throughout the specification are not necessarily all referring to the same embodiment.
• It is to be appreciated that the use of any of the following “/”, “and/or”, and “at least one of”, for example, in the cases of “A/B”, “A and/or B” and “at least one of A and B”, is intended to encompass the selection of the first listed option (A) only, or the selection of the second listed option (B) only, or the selection of both options (A and B). As a further example, in the cases of “A, B, and/or C” and “at least one of A, B, and C”, such phrasing is intended to encompass the selection of the first listed option (A) only, or the selection of the second listed option (B) only, or the selection of the third listed option (C) only, or the selection of the first and the second listed options (A and B) only, or the selection of the first and third listed options (A and C) only, or the selection of the second and third listed options (B and C) only, or the selection of all three options (A and B and C). This may be extended, as readily apparent by one of ordinary skill in this and related arts, for as many items listed.
• Turning toFIG. 1, exemplary example-based data pruning system using intra-frame patch similarity is indicated generally by thereference numeral100. Thepruning system100 includes apruner105 having an output connected in signal communication with an input of avideo encoder110 and a first input of a metadata generator andencoder135. An output of the video encoder is connected in signal communication with an input of avideo decoder115 and an input of apatch library creator140. An output of thevideo decoder115 is connected in signal communication with a first input of arecovery device120. An output of thepatch library creator130 is connected in signal communication with a second input of therecovery device120. An output of the metadata generator andencoder135 is connected in signal communication with an input of ametadata decoder125. An output of themetadata decoder125 is connected in signal communication with a third input of therecovery device120. An output of thepatch library creator140 is connected in signal communication with a second input of the metadata generator andencoder135. An output of a clustering device andpatch library creator145 is connected in signal communication with a second input of thepruner105. An input of thepruner105 and an input of the clustering device andpatch library creator145 are available as inputs to thepruning system100, for receiving input video. An output of the recovery device is available as an output of thepruning system100, for outputting video.
• Turning toFIG. 2, an exemplary video encoder to which the present principles may be applied is indicated generally by thereference numeral200. Thevideo encoder200 includes aframe ordering buffer210 having an output in signal communication with a non-inverting input of a combiner285. An output of the combiner285 is connected in signal communication with a first input of a transformer andquantizer225. An output of the transformer andquantizer225 is connected in signal communication with a first input of anentropy coder245 and a first input of an inverse transformer and inverse quantizer250. An output of theentropy coder245 is connected in signal communication with a first non-inverting input of a combiner290. An output of the combiner190 is connected in signal communication with a first input of anoutput buffer235.
• A first output of anencoder controller205 is connected in signal communication with a second input of theframe ordering buffer210, a second input of the inverse transformer and inverse quantizer250, an input of a picture-type decision module215, a first input of a macroblock-type (MB-type)decision module220, a second input of an intra prediction module260, a second input of adeblocking filter265, a first input of amotion compensator270, a first input of a motion estimator275, and a second input of areference picture buffer280.
• A second output of theencoder controller205 is connected in signal communication with a first input of a Supplemental Enhancement Information (SEI) inserter230, a second input of the transformer andquantizer225, a second input of theentropy coder245, a second input of theoutput buffer235, and an input of the Sequence Parameter Set (SPS) and Picture Parameter Set (PPS)inserter240.
• An output of the SEI inserter230 is connected in signal communication with a second non-inverting input of the combiner290.
• A first output of the picture-type decision module215 is connected in signal communication with a third input of theframe ordering buffer210. A second output of the picture-type decision module215 is connected in signal communication with a second input of a macroblock-type decision module220.
• An output of the Sequence Parameter Set (SPS) and Picture Parameter Set (PPS)inserter240 is connected in signal communication with a third non-inverting input of the combiner290.
• An output of the inverse quantizer and inverse transformer250 is connected in signal communication with a first non-inverting input of acombiner219. An output of thecombiner219 is connected in signal communication with a first input of the intra prediction module260 and a first input of thedeblocking filter265. An output of thedeblocking filter265 is connected in signal communication with a first input of areference picture buffer280. An output of thereference picture buffer280 is connected in signal communication with a second input of the motion estimator275 and a third input of themotion compensator270. A first output of the motion estimator275 is connected in signal communication with a second input of themotion compensator270. A second output of the motion estimator275 is connected in signal communication with a third input of theentropy coder245.
• An output of themotion compensator270 is connected in signal communication with a first input of aswitch297. An output of the intra prediction module260 is connected in signal communication with a second input of theswitch297. An output of the macroblock-type decision module220 is connected in signal communication with a third input of theswitch297. The third input of theswitch297 determines whether or not the “data” input of the switch (as compared to the control input, i.e., the third input) is to be provided by themotion compensator270 or the intra prediction module260. The output of theswitch297 is connected in signal communication with a second non-inverting input of thecombiner219 and an inverting input of the combiner285.
• A first input of theframe ordering buffer210 and an input of theencoder controller205 are available as inputs of theencoder200, for receiving an input picture. Moreover, a second input of the Supplemental Enhancement Information (SEI) inserter230 is available as an input of theencoder200, for receiving metadata. An output of theoutput buffer235 is available as an output of theencoder200, for outputting a bitstream.
• Turning toFIG. 3, an exemplary video decoder to which the present principles may be applied is indicated generally by thereference numeral300. Thevideo decoder300 includes aninput buffer310 having an output connected in signal communication with a first input of an entropy decoder345. A first output of the entropy decoder345 is connected in signal communication with a first input of an inverse transformer andinverse quantizer350. An output of the inverse transformer andinverse quantizer350 is connected in signal communication with a second non-inverting input of a combiner325. An output of the combiner325 is connected in signal communication with a second input of adeblocking filter365 and a first input of an intra prediction module360. A second output of thedeblocking filter365 is connected in signal communication with a first input of areference picture buffer380. An output of thereference picture buffer380 is connected in signal communication with a second input of amotion compensator370.
• A second output of the entropy decoder345 is connected in signal communication with a third input of themotion compensator370, a first input of thedeblocking filter365, and a third input of the intra predictor360. A third output of the entropy decoder345 is connected in signal communication with an input of adecoder controller305. A first output of thedecoder controller305 is connected in signal communication with a second input of the entropy decoder345. A second output of thedecoder controller305 is connected in signal communication with a second input of the inverse transformer andinverse quantizer350. A third output of thedecoder controller305 is connected in signal communication with a third input of thedeblocking filter365. A fourth output of thedecoder controller305 is connected in signal communication with a second input of the intra prediction module360, a first input of themotion compensator370, and a second input of thereference picture buffer380.
• An output of themotion compensator370 is connected in signal communication with a first input of aswitch397. An output of the intra prediction module360 is connected in signal communication with a second input of theswitch397. An output of theswitch397 is connected in signal communication with a first non-inverting input of the combiner325.
• An input of theinput buffer310 is available as an input of thedecoder300, for receiving an input bitstream. A first output of thedeblocking filter365 is available as an output of thedecoder300, for outputting an output picture.
• As noted above, the present principles are directed to methods and apparatus for example-based data pruning using intra-frame patch similarity.
• In accordance with the present principles, this application discloses a new approach that takes advantage of patch similarity within a video frame. The patch similarity within an image happens in many real-world pictures where there are repetitive textures or patterns in the pictures such as, for example, a picture with wall papers as the background. The within-picture patch similarity is discovered by a clustering algorithm, and a patch library is created for pruning and recovery. However, since the same frame is used for both creating patch library and for pruning/recovery, the patch dependency problems have to be resolved in order to ensure artifact-free recovery.
• The present principles provide an improvement of our previous approach by training the patch library at the decoder side using previously sent frames or existing frames, rather than sending the patch library through one or more communication channels. Moreover, the data pruning is realized by replacing some blocks in the input frames with flat regions to create “mixed-resolution” frames.
• As noted above, in the MPEG-4 AVC Standard, intra-frame block prediction is realized by block prediction from the neighboring blocks. However, long-range similarity of non-neighboring blocks is not exploited to increase compression efficiency. Advantageously, the present principles provide a method for pruning an input video so that the input video can be more efficiently encoded by a video encoder. The present principles take advantage of the similarity of image patches within a video frame to further increase the compression efficiency.
• In accordance with the present principles, intra-frame patch similarity is used to train an example patch library and prune a video and recover the pruned video. Error-bounded clustering (the modified K-means clustering) is used for efficient patch searching in the library. To improve compression efficiency, a mixed-resolution data pruning scheme is used, where blocks are replaced by flat blocks to reduce the high-frequency signal.
• The present principles may involve the use of patch signature matching, a matching rank list, and rank number encoding to increase the efficiency of metadata (best-match patch position in library) encoding. Moreover, a method is disclosed for encoding the block coordinates using the flat block identification based on color variation.
• Referring back toFIG. 1, one difference between the present principles and the aforementioned first approach is that the input video frame for pruning is also used for patch library creation. In thepruning system100, the encoder-side processing component can be considered to include two parts, namely a patch library creation part and a pruning part. For the encoder side, two patch libraries are generated at the encoder side, one patch library from the original frame, the other patch library from the reconstructed frame (i.e., a pruned, encoded and then decoded frame). The latter is exactly the same as the patch library created at the encoder side in that they use exactly the same frame (i.e., the reconstructed frame) to generate the patch libraries. At the encoder side, the patch library created using the original frame is used to prune the blocks, whereas the patch library created using the reconstructed frame is used to encode metadata. The reason of using a patch library created from the reconstructed frame is to make sure the patch libraries for encoding and decoding metadata are identical at the encoder and decoder side. For the patch library created using the original frames, a clustering algorithm is performed to group the patches so that the patch search process during pruning can be efficiently carried out. Pruning is a process to modify the source video using the patch library so that less bits are sent to the decoder side. Pruning is realized by dividing a video frame into blocks, and replacing some of the blocks with flat blocks. The pruned frame is then taken as the input for a video encoder. An example video encoder to which the present principles may be applied is shown inFIG. 2 described above.
• Referring back toFIG. 1, the decoder-side processing component of thepruning system100 can also be considered to include two parts, namely a patch library creation part and a recovery part. Patch library creation is a process to create a patch library that is exactly the same as the library used for pruning at the encoder side. This is ensured by using exactly the same frame (i.e., the reconstructed pruned frame) for patch library creation. The recovery component is a process to recover the pruned content in the decoded pruned frames sent from the encoder side. An example video encoder to which the present principles may be applied is shown inFIG. 3 described above.
Patch Library Creation
• Turning toFIG. 4, an exemplary first portion for performing encoder side processing in an example-based data pruning system using intra-frame patch similarity is indicated generally by thereference numeral400. Thefirst portion400 includes adivider410 having an output in signal communication with an input of aclustering device420. An input of the divider is available as an input to thefirst portion400, for receiving the first frame in a GOP. An output of theclustering device420 is available as an output of thefirst portion400, for outputting clusters and a patch library.
• Turning toFIG. 5, an exemplary method for clustering and patch library creation is indicated generally by thereference numeral500. Atstep505, a training video frame is input. Atstep510, the training video frame is divided (by divider410) into overlapping blocks. Atstep515, blocks without high-frequency details are removed (by the clustering device420). Atstep520, the blocks are clustered (by the clustering device420). Atstep525, clusters and a patch library are output.
• The patch library is a pool of high resolution patches that can be used to recover pruned image blocks. Turning toFIG. 6, an exemplary patch library and corresponding clusters are indicated generally by thereference numeral600. The patch library is specifically indicated by thereference numeral610, and includes asignature portion611 and a highresolution patch portion612. For the encoder side processing, two patch libraries are generated, one patch library for pruning, the other patch library for metadata encoding. The patch library for pruning is generated using the original frame, whereas the patch library for metadata encoding is generated using the reconstructed frame. For the patch library for pruning, the patches in the library are grouped into clusters so that the pruning search process can be efficiently performed. The video frames used for library creation are divided into overlapping blocks to form a training data set. The training data is first cleaned up by removing all blocks that do not include high-frequency details. A modified K-means clustering algorithm is used to group the patches in the training data set into clusters. For each cluster, the cluster center is the average of the patches in the cluster, and is used for matching to an incoming query during the pruning process. The modified K-means clustering algorithm guarantees that the error between any patch within a cluster and its cluster center is smaller than a specified threshold. The modified K-means clustering algorithm could be replaced by any similar clustering algorithm which ensures the error bound in the clusters.
• To speed up computation, the horizontal and vertical dimensions of the training frames are reduced to one quarter of the original size. Also, the clustering process is performed on the patches in the downsized frames. In one exemplary embodiment, the size of the high-resolution patches is 16×16 pixels, and the size of the downsized patches is 4×4 pixels. Therefore, the downsize factor is 4. Of course, other sizes can be used, while maintaining the spirit of the present principles.
• For the patch library for metadata encoding, the clustering process and clean-up process are not performed, therefore it includes all possible patches from the reconstructed frame. However, for every patch in the patch library created from the original frames, its corresponding patch can be found in the patch library created from the reconstructed frame using the coordinates of the patches. This would make sure that metadata encoding can be correctly performed. For the decoder side, the same patch library without clustering is created using the same decoded video frames for metadata decoding and pruned block recovery.
• For the patch libraries created using decoded frames at both the encoder and decoder side, another process is conducted to create the signatures of the patches. The signature of a patch is a feature vector that includes the average color of the patch and the surrounding pixels of the patch. The patch signatures are used for the metadata encoding process to more efficiently encode the metadata, and used in the recovery process at the decoder side to find the best-match patch and more reliably recover the pruned content. Turning toFIG. 7, an exemplary signature vector is indicated generally by thereference numeral700. Thesignature vector700 includes anaverage color701 and surroundingpixels702.
• The metadata encoding process is described herein below. In the pruned frame, sometimes the neighboring blocks of a pruned block for recovery or metadata encoding are also pruned. Then the set of surrounding pixels used as the signature for search in the patch library only includes the pixels from the non-pruned blocks. If all the neighboring blocks are pruned, then only theaverage color701 is used as the signature. This may end up with bad patch matches since too little information is used for patch matching, that is why neighboringnon-pruned pixels702 are important.
Pruning Process
• Similar to standard video encoding algorithms, the input video frames are divided into Group of Pictures (GOP). The pruning process is conducted on the first frame of a GOP. The pruning result is propagated to the rest of the frames in the GOP afterwards.
Pruning Process for the First Frame in a GOP
• Turning toFIG. 8, an exemplary second portion for performing encoder side processing in an example-based data pruning system using intra-frame patch similarity is indicated generally by thereference numeral800. Thesecond portion800 includes adivider805 having an output in signal communication with an input of apatch library searcher810. An output of thepatch library searcher810 is connected in signal communication with an input of avideo encoder815, a first input of a metadata generator830, and a first input of ametadata encoder825. An output of the metadata generator830 is connected in signal communication with a second input of themetadata encoder825. A first output of thevideo encoder815 is connected in signal communication with a second input of the metadata generator830. An input of thedivider805 is available as an input of thesecond portion800, for receiving an input frame. An output of thevideo encoder815 is available as an output of thesecond portion800, for outputting an encoded video frame. An output of themetadata encoder825 is available as an output of thesecond portion800, for outputting encoded metadata.
• Turning toFIG. 9, an exemplary method for pruning a video frame is indicated generally by thereference numeral900. Atstep905, a video frame is input. Atstep910, the video frame is divided into non-overlapping blocks. Atstep915, a loop is performed for each block. Atstep920, a search is performed in the patch library. Atstep925, it is determined whether or not a patch has been found. If so, then the method proceeds to step930. Otherwise, the method returns to step915. Atstep930, the block is pruned. Atstep935, it is determined whether or not all blocks have been finished. If so, then the method proceeds to step940. Otherwise, the method returns to step915. Atstep940, the pruned frame and corresponding metadata are output.
• Thus, the input frame is first divided into non-overlapping blocks perstep910. The size of the block is the same as the size of the macroblock used in the standard compression algorithms, in our current implementation, 16×16 pixels. A search process then is followed to find the best-match patch in the patch library perstep920. This search process is illustrated inFIG. 10. Turning toFIG. 10, a patch search process performing during pruning is indicated generally by thereference numeral1000. Thepatch search process1000 involves apatch library1010 which, in turn, includes asignature portion1011 and a highresolution patch portion1012. First, the block is matched with the centers of the clusters by calculating the Euclidean distance, and finding the top K matched clusters. Currently, K is determined empirically. In principle, K is determined by the error bound of the clusters. Of course, other approaches to calculating K may also be used in accordance with the teachings of the present principles. After the candidate clusters are indentified, the search process is conducted within the clusters until the best-match patch is found in the clusters. If the difference between the best-match patch and the query block is sufficiently small, the block would be pruned. Otherwise, the block will be kept intact.
• The preceding approach is different from the aforementioned first approach, since in the first approach the patches in the input frames are used to create patch library and recover the pruned blocks, thus resulting in block dependency problem, meaning that the needed patches for recovery may not be available when a block is being recovered. This problem is solved by following two solutions provided by the present principles: pruning with causal recovery; and pruning using a dependency graph.
a) Pruning for Causal Recovery
• In the case of pruning with causal recovery, for a pruned block, the search process (also the recovery process at the decoder side) will only look at the patches preceding the pruned block in the coordinates in the patch library. Turning toFIG. 11, a block search area in a process of pruning for causal recovery is indicated generally by thereference numeral1100. Theblock search area1100 includes anexample patch1110 and acandidate block1120. The patches preceding the pruning block (the candidate block1120) are within the shaded area inFIG. 11. In recovery, by the time the prunedblock1120 is being recovered, all blocks within the shaded area have been recovered, therefore all patches within the shaded area should be available.
• b) Pruning with Dependency Graph
• The limitation of the above approach is that there may be no example patches available for the blocks at the top of a frame. This problem may be solved by a full-frame patch search with the help of a patch dependency graph. The patch dependency graph is a directed acyclic graph (DAG), where each node of the graph represents a candidate block for pruning, and each edge represents the dependency of the blocks. Turning toFIG. 12, an exemplary block dependency graph is indicated generally by thereference numeral1200. InFIG. 12, each circle is a node of the dependency graph, which represents a block in a video frame. The arrows (edges of the graph) between the circles indicate the dependencies of the blocks. If an arrow starts from one circle and points to another, then the circle that the arrow starts from is dependent on the circle that the arrow points to. If a circle does not have any arrow pointing to any other circle, the corresponding block is an unpruned block that does not need recovery after decoding. Thedependency graph1200 is created during the block search process. For a candidate block, after the block search process, if the best-match patch is found, then an edge that points from the candidate block to the blocks overlapping with the best-match patches would be created in the dependency graph. After the search process is done for all the candidate blocks, a process is carried out to obtain the recovery sequence of the pruned blocks. Turning toFIG. 13, an exemplary method for obtaining a recovery sequence is indicated generally by thereference numeral1300. Atstep1305, a block dependency graph is input. Atstep1310, end nodes are found in the block dependency graph. Atstep1315, block IDs are saved and end nodes are removed. Atstep1320, it is determined whether or not any end nodes remain. If so, then the method proceeds to step1325. Otherwise, the method returns to step1310. Atstep1325, it is determined whether or not the graph is empty. If so, then the method proceeds to step1330. Otherwise, the method proceeds to step1335. Atstep1330, the block IDs are output and saved as a recovery sequence. Atstep1335, a node with the maximum indegree is found. At step1340, the node is removed from the block dependency graph.
• The method for obtaining a recovery sequence is further described as follows:
• 1. Find all end nodes (i.e., the nodes that do not depend on other nodes) in the graph and save the corresponding block coordinates (here the block coordinates are represented as the block IDs as shown inFIG. 12) of all the end nodes. After the block IDs are saved, the corresponding nodes are removed from the dependency graph. The procedure repeats until all end nodes are removed. If the dependency graph is empty, the algorithm ends, otherwise the algorithm goes to step (2).
• 2. Find a node in the graph with maximum indegree, i.e., a node corresponding to the block upon which depends a maximum number of other blocks. Remove the node in the graph, do not save the IDs of the block (i.e., the block will not be pruned). Repeat this procedure until there are new end nodes emerging in the graph. Then the algorithm goes back to step (1). The block is not pruned because the block cannot be recovered using the available pixels (decoded pixels and recovered pixels) in the frame. On the other hand, other blocks may depend on this block to recover. Therefore, in order to prune maximum number of blocks, the block upon which depends the maximum number of other blocks is found. After the block is kept unpruned (i.e., removed from the graph), there may be new end nodes emerging, and then step (1) can be used to prune the block again.
• Turning toFIG. 14, an exemplary evolution of the dependency graph using the pruning algorithm is indicated generally by thereference numeral1400. The evolution involves the graph before pruning1410, the graph after step (1) is performed (1420), and the graph after step (2) is performed (1430).
• By using the above algorithm, a block recovery sequence which ensures that the best-matching patch is available will be obtained when a corresponding block is being recovered during the recovery process.
• After the blocks are identified for pruning, a process is conducted to prune the block. There could be different pruning strategies. For example, replacing the high-resolution blocks with low-resolution blocks may be one strategy that is used. However, it is discovered that it is difficult for this approach to achieve significant compression efficiency gain. Therefore, in the current system, a high-resolution block is simply replaced with a flat block, in which all pixels have the same color value, which is the average of the color values of the pixels within the original block. The block replacement process creates a video frame where some parts of the frame have high-resolution and some parts have low-resolution; therefore, such a frame is called as mixed-resolution frame. Turning toFIG. 15, an exemplary mixed-resolution frame is indicated generally by thereference numeral1500. Our experiments show that such flat-block replacement scheme is quite effective to gain compression efficiency. The flat block replacement scheme could be replaced by a low-resolution block replacement scheme, where the block for pruning is replaced by its low-resolution version.
Metadata Encoding and Decoding
• Metadata encoding includes two components (seeFIG. 16), one for encoding pruned block IDs (seeFIG. 17), the other for encoding patch index (seeFIG. 18), which are the results of searching patch library for each block during the pruning process.
• Turning toFIG. 16, an exemplary method for encoding metadata is indicated generally by thereference numeral1600. Atstep1605, a decoded pruned video frame, pruned block IDs, and a patch index for each block are input. Atstep1610, pruned block IDs are encoded. Atstep1615, the patch index is encoded. Atstep1620, the encoded metadata is output.
• Turning toFIG. 17, an example method for encoding pruned block IDs is indicated generally by thereference numeral1700. Atstep1705, a pruned frame and pruned block IDs are input. Atstep1710, a low-resolution block identification is performed. Atstep1720, it is determined whether or not there are any misses. If so, then the method proceeds to step1725. Otherwise, the method proceeds to step1715. Atstep1725, it is determined whether or not the number of false positives is more than the number of pruned blocks. If so, then the method proceeds to step1630. Otherwise, control proceeds to step1735. Atstep1730, the pruned block sequence is used, and a flag is set equal to zero. Atstep1740, a differentiation is performed. Atstep1745, lossless encoding is performed. Atstep1750, the encoded metadata is output. Atstep1715, a threshold is adjusted. Atstep1735, the false positive sequence is used, and the flag is set equal to one.
• Turning toFIG. 18, an exemplary method for encoding a patch index is indicated generally by thereference numeral1800. Atstep1805, a decoded pruned video frame and a patch index for each block are input. Atstep1810, a loop is performed for each pruned block. Atstep1815, a signature is obtained. Atstep1820, the distances to the patches in the patch library are calculated. Atstep1825, the patches are sorted to obtain a rank list. Atstep1830, the rank number is obtained. Atstep1835, the rank number is entropy coded. Atstep1840, it is determined whether or not all blocks are finished (being processed). If so, then the method proceeds to step1845. Otherwise, the method returns to step1810. Atstep1845, the encoded patch index is output.
• During the pruning process, for each block, the system would search the best match patch in the patch library and output a patch index in the patch library for a found patch if the distortion is less than a threshold. Each patch is associated with its signature (i.e., its color plus surrounding pixels in the decoded frames). During the recovery process in the decoder side processing, the color of the pruned block and its surrounding pixels are used as a signature to find the correct high-resolution patch in the library.
• However, due to noise, the search process using the signature is not reliable, and metadata is needed to assist the recovery process to ensure reliability. Therefore, after the pruning process, the system will proceed to generate metadata for assisting recovery. For each pruned block, the search process described above already identifies the corresponding patches in the library. The metadata encoding component will simulate the recovery process to encode the metadata. A new patch library will be created using the decoded pruned frame to ensure the patch library is exactly the same as that in the decoder side. The frame is divided into overlapping patches and signatures are created for the patches. During the recovery simulation process, the patch library has to be dynamically updated because some pruned blocks will be recovered during the process. The process is illustrated inFIG. 17. Thus, referring back toFIG. 18, for each pruned block, its query signature (the average color of the pruned block plus the surrounding pixels) will be used to match the signatures of the patches in the library. For each block, the distances (e.g., Euclidean, although, of course, other distance metrics may be used) between the query vector and the signatures of the patches in the library are calculated. The patches are sorted according to the distances, resulting in a rank list. In the ideal case, the best-match high-resolution patch should be at the top of the rank list. However, due to the noise caused by arithmetic rounding and compression, the best-match patch is often not the first one in the rank list. Presume that the correct patch is the n^thpatch in the rank list. The rank number n will be saved as the metadata for the block. It should be noted that, in the most cases, n is 1 or a very small number because the best-match patch is close to the top in the rank list, therefore the entropy of this random number is significantly smaller than the index of the best-match patch in the library, which should be a uniform distribution having maximum entropy. Therefore, the rank number can be more efficiently encoded by entropy coding. The rank numbers of all the pruned blocks form a rank number sequence as part of the metadata sent to the decoder side. It is observed from the actual experiments that the distribution of the rank numbers is close to a geometric distribution, therefore currently the Golomb code is used for further encoding the rank number sequence, since Golomb code is optimal for a random number having geometric distribution. Of course, other types of codes may also be used in accordance with the teachings of the present principles, while maintaining the spirit of the present principles.
• For decoding (seeFIG. 19), the decoder side should have the exactly the same patch library as the encoder, as the signature of the pruned block will be used to match with the signatures in the patch library and get a rank list (the sorted patch library). The rank number is used to retrieve the correct patch from the sorted patch library. If the patch library is created from previous frames, in order to ensure the encoder and decoder side have exactly the same patch library, the metadata encoding process at the encoder side should also use the decoded frames from the video decoder because only the decoded frames are available at the decoder side.
• Turning toFIG. 19, an exemplary method for decoding a patch index is indicated generally by thereference numeral1900. Atstep1905, a decoded pruned video frame, an encoded patch index, and pruned block IDs are input. Atstep1910, a loop is performed for each pruned block. Atstep1915, a signature is obtained. Atstep1920, the distances to the patches in the patch library are calculated. Atstep1925, the patches are sorted to obtain a rank list. Atstep1930, the encoded rank number is entropy decoded. Atstep1935, the patch index is retrieved from the patch library using the rank number. Atstep1940, it is determined whether or not all blocks are finished (being processed). If so, then the method proceeds to step1945. Otherwise, the method returns to step1910. Atstep1945, the decoded patch index is output.
• Besides the rank number metadata, it is necessary to send the locations of the pruned blocks to the decoder side. This is done by block ID encoding (seeFIG. 17). One simple way is to just send a block ID sequence to the decoder side. The ID of a block indicates the coordinate of the block on the frame. Turning toFIG. 20, an exemplary block ID is indicated generally by thereference numeral2000. However, it is also possible to more efficiently encode the ID sequence of the pruned blocks. Since the pruned blocks are flat and do not include any high-frequency components, it is possible to detect the pruned blocks by calculating the color variation within the blocks. If the color variation is smaller than a threshold, then the block is identified as a pruned block. However, since such an identification process is not reliable due to noise caused by compression, metadata is needed to facilitate the identification process. First, the variance threshold is determined by starting from a high threshold value. The algorithm then slowly decrease the variance threshold such that all pruned blocks can be identified by the identification procedure, but false positive blocks may be present in the identified results. Afterwards, if the number of the false positives is larger than that of the pruned blocks, the IDs of the pruned blocks are saved and sent to decoder, otherwise the IDs of the false positives would be sent to the decoder side. The variance threshold for identifying flat blocks is also sent to the decoder side for running the same identification procedure. For the pruning method with causal recovery as described above, the order of the block IDs does not matter, so the ID sequence can be sorted so that the numbers are increasing.
• To further reduce redundancy, it is possible to have a differential coding scheme to compute the difference between an ID number to its previous ID number, and encode the difference sequence. For example, assuming the ID sequence is 3, 4, 5, 8, 13, 14, the differentiated sequence becomes 3, 1, 1, 3, 5, 1. The differentiation process makes the numbers closer to 1, therefore resulting in a number distribution with smaller entropy. The differentiated sequence then can be further encoded with entropy coding (e.g., Golomb code in our current implementation). Thus, the format of the final metadata is shown as follows:
• 
• where flag is a signaling flag to indicate whether or not the block ID sequence is a false positive ID sequence, the threshold is the variance threshold for flat block identification, the encoded block ID sequence is the encoded bit stream of the pruned block IDs or the false positive block IDs, and the encoded rank number sequence is the encoded bit stream of the rank numbers used for block recovery.
Pruning Process for the Rest of the Frames in a GOP
• For the rest of the frames in a GOP, some of the blocks in the frames will be also replaced by flat blocks. The positions of the pruned blocks in the first frame can be propagated to the rest of the frames by motion tracking. Different strategies are tried to propagate the positions of the pruned blocks. One approach is to track the pruned blocks across frames by block matching, and prune the corresponding blocks in the subsequent frames (i.e., replace the tracked blocks with flat blocks). However, this approach does not result in good compression efficiency gain because, in general, the boundaries of the tracked blocks do not align with the coding macro blocks. As a result, the boundaries of the tracked blocks create a high frequency signal in the macro blocks. Therefore, a simpler alternative approach is currently used to set all the block positions for the subsequent frames to the same positions as the first frame. Namely, all the pruned blocks in the subsequent frames are collocated with the pruned blocks in the first frame. As a result, all of the pruned blocks for the subsequent frames are aligned with macro block positions.
• However, this approach would not work well if there is motion in the pruned blocks. Therefore, one solution to solve the problem is to detect calculate the motion intensity of the block (seeFIG. 21). Turning toFIG. 21, an exemplary method for pruning sequent frames is indicated generally by thereference numeral2100. Atstep2105, a video frame and pruned block IDs are input. Atstep2110, collocated blocks are pruned. Atstep2115, a loop is performed for each block. Atstep2120, a motion vector is calculated to the previous frame. Atstep2125, the motion vectors are saved as metadata. Atstep2130, it is determined whether or not all blocks are finished (being processed). If so, then the method proceeds to step2135. Otherwise, the method returns to step2115.
• If the motion intensity is larger than a threshold, the block would not be pruned. Another more sophisticated solution, which is our current implementation, is to calculate the motion vectors of the pruned blocks in the original video by finding the corresponding block in the previous frame (seeFIG. 22). Turning toFIG. 22, an exemplary motion vector for a pruned block is indicated generally by thereference numeral2200. Themotion vector2200 relates to a pruned block in an i-th frame and a co-located block in a (i−1)-th frame. The motion vectors of the pruned blocks would be sent to the decoder side for a recovery purpose. Since the previous frame would already have been completely recovered, the pruned blocks in the current frame can be recovered using the motion vectors. To avoid artifacts, if the difference between the block in the current frame and the corresponding patch block calculated by motion estimation in the previous frame is too large, then the block in the current frame would not be pruned. Furthermore, subpixel motion estimation is currently used to make motion vector based recovery more accurate. Our experiments show that the resultant visual quality using subpixel based motion vector estimation is much better than that using regular pixel based motion vector estimation.
Recovery Process
• The recovery process takes place at the decoder side. The patch library is created before the recovery process by obtaining all the overlapping patches and creating the signatures using the first decoded frame in the GOP. However, different from the aforementioned first approach, the patch library has to be dynamically updated during the recovery process, because the pruned blocks in the frame will be replaced with the recovered blocks during the recovery process.
• For the first frame in a GOP, the recovery process starts with decoding the metadata (seeFIG. 23), including decoding the block ID sequence (seeFIG. 24) and the rank order sequence (seeFIG. 23). Turning toFIG. 23, an exemplary method for decoding metadata is indicated generally by thereference numeral2300. Atstep2305, encoded metadata is input. Atstep2310, pruned block IDS are decoded. Atstep2315, a patch index is decoded. Atstep2320, decoded metadata is output.
• Turning toFIG. 24, an exemplary method for decoding pruned block IDs is indicated generally by thereference numeral2400. Atstep2405, encoded metadata is input. Atstep2410, lossless decoding is performed. Atstep2415, reverse differentiation is performed. Atstep2420, it is determined whether or not a flag is equal to zero. If so, then the method proceeds to step2425. Otherwise, the method proceeds to step2430. Atstep2425, block IDs are output. Atstep2430, a low resolution block identification is performed. Atstep2435, false positives are removed. Atstep2440, block IDs are output.
• After the block ID sequence is available, for each pruned block, the average color and surrounding pixels of this block will be taken as the signature to match with the signatures in the patch library. However, if the neighboring blocks of the block for recovery are also pruned, then the set of surrounding pixels used as the signature for search only includes the pixels from the non-pruned blocks. If all the neighboring blocks are pruned, then only the average color is used as the signature. The matching process is realized by calculating the Euclidean distances between the signature of the query block and those of the patches in the library. After all the distances are calculated, the list is sorted according to the distances, resulting in a rank list. The rank number corresponding to the pruned block then is used to retrieve the correct high-resolution block from the rank list.
• Turning toFIG. 25, an exemplary apparatus for performing decoder side processing for example-based data pruning using intra-frame patch similarity is indicated generally by thereference numeral2500. Theapparatus2500 includes adivider2505 having an output connected in signal communication with a first input of a search patch library andreplacement block device2510. An output of ametadata decoder2515 is connected in signal communication with a second input of the search patch library andreplacement block device2510. An input of thedivider2505 is available as an input of theapparatus2500, for receiving pruned video. An input of themetadata decoder2515 is available as an input of theapparatus2500, for receiving encoded metadata. An output of the search patch library andreplacement block device2510 is available as an output of the apparatus, for outputting recovered video.
• Turning toFIG. 26, an exemplary method for recovering a pruned frame is indicated generally by thereference numeral2600. Atstep2605, a pruned frame and corresponding metadata are input. Atstep2610, the pruned frame is divided into non-overlapping blocks. Atstep2615, a loop is performed for each block. Atstep2620, it is determined whether or not the current block is a pruned block. If so, then the method proceeds to step2625. Otherwise, the method returns to step2615. Atstep2625, a patch is found in the library. Atstep2630, a current block is replaced with the found patch. Atstep2635, it is determined whether or not all blocks are finished (being processed). If so, then the method proceeds to step2640. Otherwise, the method returns to step2615. Atstep2640, the recovered frame is output.
• It is to be appreciated that the block recovery using example patches can be replaced by traditional inpainting and texture synthesis based methods.
• Note that for the pruning scheme with dependency graph as described above, the recovery process has to follow the order of the block IDs in the ID sequence so that whenever a block is being recovered, its corresponding patch is available in the patch library. Furthermore, after each block is recovered, the patch library has to be updated, i.e., the patches overlapping with the block have to be replaced with new patches and the signatures for those patches and their neighbors have to be recalculated.
• For the rest of the frames in a GOP, for each pruned block, if the motion vector is not available, the content of the block can be copied from the co-located block in the previous frame. If the motion vector is available, then it is possible to use the motion vector to find the corresponding block in the previous frame, and copy the corresponding block to fill the pruned block. Turning toFIG. 27, an exemplary method for recovering subsequent frames is indicated generally by thereference numeral2700. Atstep2705, a video frame and pruned block IDs are input. Atstep2710, a loop is performed for each block. Atstep2715, a motion vector is used to find the patch in the previous frame. Atstep2720, the found patch is used to replace the pruned block. Atstep2725, it is determined whether or not all blocks are finished (being processed). If so, then the method proceeds to step2730. Otherwise, the method returns to step2710.
• Block artifacts may be visible since the recovery process is block-based. A deblocking filter, such as the in-loop deblocking filter used in the MPEG-4 AVC Standard encoder, can be applied to reduce the block artifacts.
• These and other features and advantages of the present principles may be readily ascertained by one of ordinary skill in the pertinent art based on the teachings herein. It is to be understood that the teachings of the present principles may be implemented in various forms of hardware, software, firmware, special purpose processors, or combinations thereof.
• Most preferably, the teachings of the present principles are implemented as a combination of hardware and software. Moreover, the software may be implemented as an application program tangibly embodied on a program storage unit. The application program may be uploaded to, and executed by, a machine comprising any suitable architecture. Preferably, the machine is implemented on a computer platform having hardware such as one or more central processing units (“CPU”), a random access memory (“RAM”), and input/output (“I/O”) interfaces. The computer platform may also include an operating system and microinstruction code. The various processes and functions described herein may be either part of the microinstruction code or part of the application program, or any combination thereof, which may be executed by a CPU. In addition, various other peripheral units may be connected to the computer platform such as an additional data storage unit and a printing unit.
• It is to be further understood that, because some of the constituent system components and methods depicted in the accompanying drawings are preferably implemented in software, the actual connections between the system components or the process function blocks may differ depending upon the manner in which the present principles are programmed. Given the teachings herein, one of ordinary skill in the pertinent art will be able to contemplate these and similar implementations or configurations of the present principles.
• Although the illustrative embodiments have been described herein with reference to the accompanying drawings, it is to be understood that the present principles is not limited to those precise embodiments, and that various changes and modifications may be effected therein by one of ordinary skill in the pertinent art without departing from the scope or spirit of the present principles. All such changes and modifications are intended to be included within the scope of the present principles as set forth in the appended claims.

Claims (30)

1. An apparatus for encoding a picture in a video sequence, comprising:

a patch library creator for creating a first patch library from an original version of said picture and a second patch library from a reconstructed version of said picture, each of said first patch library and said second patch library including a plurality of high resolution replacement patches for replacing one or more pruned blocks during a recovery of a pruned version of said picture; and

a pruner for generating said pruned version of said picture from said first patch library;

a metadata generator for generating metadata from said second patch library, said metadata for recovering said pruned version of said picture; and

an encoder for encoding said pruned version of said picture and said metadata,

wherein said first patch library includes a plurality of patch clusters, and said pruned version of said picture is generated by dividing said original version of said picture into a plurality of overlapping blocks, searching for candidate patch clusters from among said plurality of patch clusters for each of said plurality of overlapping blocks based on respective distance metrics from each of said plurality of overlapping blocks to respective centers of each of said plurality of patch clusters, identifying a best matching patch from said candidate patch clusters based on one or more criterion, and pruning a corresponding one of said plurality of overlapping blocks to obtain a pruned block there for when a difference between said corresponding one of said plurality of overlapping blocks and said best matching patch is less than a threshold difference, and

wherein a patch dependency graph having a plurality of nodes and a plurality of edges is used for said searching, each of said plurality of nodes representing a respective one of said plurality of overlapping blocks, and each of said plurality of edges representing a respective dependency of at least said respective one of said plurality of overlapping blocks.

2. The apparatus ofclaim 1, wherein said pruned version of said picture is generated by dividing said original version of said picture into a plurality of blocks, and respectively replacing at least one of said plurality of blocks with a replacement patch, wherein all pixels in said replacement patch have one of a same color value or a low resolution.

3. The apparatus ofclaim 2, wherein said same color value is equal to an average of color values of said pixels within said at least one of said plurality of blocks.

4. The apparatus ofclaim 1, wherein said first patch library is created by dividing said original version of said picture into a plurality of overlapping blocks to form a training data set, removing any of said plurality of overlapping blocks from said training set having a high frequency component above a pre-specified threshold, and clustering remaining ones of said plurality of overlapping blocks into a plurality of clusters, wherein each of said remaining ones of said plurality of overlapping blocks form a respective one of said plurality of high resolution replacement patches.

5. The apparatus ofclaim 4, wherein a respective center of a respective one of said plurality of clusters corresponds to an average of any of said remaining ones of said plurality of overlapping blocks included in said respective one of said plurality of clusters.

6. The apparatus ofclaim 5, wherein said remaining ones of said plurality of overlapping blocks are downsized prior to said clustering to obtain a plurality of downsized overlapping blocks, said clustering is performed on said plurality downsized overlapping blocks, and said respective center of said respective one of said plurality of clusters corresponds to said average of any of said plurality of downsized overlapping blocks included in said respective one of said plurality of clusters.

7. The apparatus ofclaim 1, wherein a signature is respectively created for each of said plurality of high resolution patches included in said second patch library by generating a feature vector there for that includes an average color for a respective one of said plurality of high resolution patches.

8. The apparatus ofclaim 7, wherein said average color included in said feature vector for said respective one of said plurality of high resolution patches is further of surrounding pixels with respect to said respective one of said plurality of high resolution patches.

9. The apparatus ofclaim 1, wherein only patches preceding a co-located patch with respect to said corresponding one of said plurality of overlapping blocks are used for said searching.

10. The apparatus ofclaim 1, wherein said metadata comprises a patch index for said best matching patch when said difference between said corresponding one of said plurality of overlapping blocks and said best matching patch is less than said threshold difference, said metadata further comprising a block identifier for said pruned block.

11. A method for encoding a picture in a video sequence, comprising:

creating a first patch library from an original version of said picture and a second patch library from a reconstructed version of said picture, each of said first patch library and said second patch library including a plurality of high resolution replacement patches for replacing one or more pruned blocks during a recovery of a pruned version of said picture; and

generating said pruned version of said picture from said first patch library;

generating metadata from said second patch library, said metadata for recovering said pruned version of said picture; and

encoding said pruned version of said picture and said metadata,

wherein said first patch library includes a plurality of patch clusters, and said pruned version of said picture is generated by dividing said original version of said picture into a plurality of overlapping blocks, searching for candidate patch clusters from among said plurality of patch clusters for each of said plurality of overlapping blocks based on respective distance metrics from each of said plurality of overlapping blocks to respective centers of each of said plurality of patch clusters, identifying a best matching patch from said candidate patch clusters based on one or more criterion, and pruning a corresponding one of said plurality of overlapping blocks to obtain a pruned block there for when a difference between said corresponding one of said plurality of overlapping blocks and said best matching patch is less than a threshold difference, and

wherein a patch dependency graph having a plurality of nodes and a plurality of edges is used for said searching, each of said plurality of nodes representing a respective one of said plurality of overlapping blocks, and each of said plurality of edges representing a respective dependency of at least said respective one of said plurality of overlapping blocks.

12. The method ofclaim 11, wherein said pruned version of said picture is generated by dividing said original version of said picture into a plurality of blocks, and respectively replacing at least one of said plurality of blocks with a replacement patch, wherein all pixels in said replacement patch have one of a same color value or a low resolution.

13. The method ofclaim 12, wherein said same color value is equal to an average of color values of said pixels within said at least one of said plurality of blocks.

14. The method ofclaim 11, wherein said first patch library is created by dividing said original version of said picture into a plurality of overlapping blocks to form a training data set, removing any of said plurality of overlapping blocks from said training set having a high frequency component above a pre-specified threshold, and clustering remaining ones of said plurality of overlapping blocks into a plurality of clusters, wherein each of said remaining ones of said plurality of overlapping blocks form a respective one of said plurality of high resolution replacement patches.

15. The method ofclaim 14, wherein a respective center of a respective one of said plurality of clusters corresponds to an average of any of said remaining ones of said plurality of overlapping blocks included in said respective one of said plurality of clusters.

16. The method ofclaim 14, wherein said remaining ones of said plurality of overlapping blocks are downsized prior to said clustering to obtain a plurality of downsized overlapping blocks, said clustering is performed on said plurality of downsized overlapping blocks, and said respective center of said respective one of said plurality of clusters corresponds to said average of any of said plurality of downsized overlapping blocks included in said respective one of said plurality of clusters.

17. The method ofclaim 11, wherein a signature is respectively created for each of said plurality of high resolution patches included in said second patch library by generating a feature vector there for that includes an average color for a respective one of said plurality of high resolution patches.

18. The method ofclaim 17, wherein said average color included in said feature vector for said respective one of said plurality of high resolution patches is further of surrounding pixels with respect to said respective one of said plurality of high resolution patches.

19. The method ofclaim 11, wherein only patches preceding a co-located patch with respect to said corresponding one of said plurality of overlapping blocks are used for said searching.

20. The method ofclaim 11, wherein said metadata comprises a patch index for said best matching patch when said difference between said corresponding one of said plurality of overlapping blocks and said best matching patch is less than said threshold difference, said metadata further comprising a block identifier for said pruned block.

21. An apparatus for encoding a picture in a video sequence, comprising:

means for creating a first patch library from an original version of said picture and a second patch library from a reconstructed version of said picture, each of said first patch library and said second patch library including a plurality of high resolution replacement patches for replacing one or more pruned blocks during a recovery of a pruned version of said picture;

means for generating said pruned version of said picture from said first patch library;

means for generating metadata from said second patch library, said metadata for recovering said pruned version of said picture; and

means for encoding said pruned version of said picture and said metadata,

wherein a patch dependency graph having a plurality of nodes and a plurality of edges is used for said searching, each of said plurality of nodes representing a respective one of said plurality of overlapping blocks, and each of said plurality of edges representing a respective dependency of at least said respective one of said plurality of overlapping blocks,

wherein said first patch library includes a plurality of patch clusters, and said pruned version of said picture is generated by dividing said original version of said picture into a plurality of overlapping blocks, searching for candidate patch clusters from among said plurality of patch clusters for each of said plurality of overlapping blocks based on respective distance metrics from each of said plurality of overlapping blocks to respective centers of each of said plurality of patch clusters, identifying a best matching patch from said candidate patch clusters based on one or more criterion, and pruning a corresponding one of said plurality of overlapping blocks to obtain a pruned block there for when a difference between said corresponding one of said plurality of overlapping blocks and said best matching patch is less than a threshold difference, and

wherein a patch dependency graph having a plurality of nodes and a plurality of edges is used for said searching, each of said plurality of nodes representing a respective one of said plurality of overlapping blocks, and each of said plurality of edges representing a respective dependency of at least said respective one of said plurality of overlapping blocks.

22. The apparatus ofclaim 21, wherein said pruned version of said picture is generated by dividing said original version of said picture into a plurality of blocks, and respectively replacing at least one of said plurality of blocks with a replacement patch, wherein all pixels in said replacement patch have one of a same color value or a low resolution.

23. The apparatus ofclaim 22, wherein said same color value is equal to an average of color values of said pixels within said at least one of said plurality of blocks.

24. The apparatus ofclaim 21, wherein said first patch library is created by dividing said original version of said picture into a plurality of overlapping blocks to form a training data set, removing any of said plurality of overlapping blocks from said training set having a high frequency component above a pre-specified threshold, and clustering remaining ones of said plurality of overlapping blocks into a plurality of clusters, wherein each of said remaining ones of said plurality of overlapping blocks form a respective one of said plurality of high resolution replacement patches.

25. The apparatus ofclaim 24, wherein a respective center of a respective one of said plurality of clusters corresponds to an average of any of said remaining ones of said plurality of overlapping blocks included in said respective one of said plurality of clusters.

26. The apparatus ofclaim 25, wherein said remaining ones of said plurality of overlapping blocks are downsized prior to said clustering to obtain a plurality of downsized overlapping blocks, said clustering is performed on said plurality of downsized overlapping blocks, and said respective center of said respective one of said plurality of clusters corresponds to said average of any of said plurality of downsized overlapping blocks included in said respective one of said plurality of clusters.

27. The apparatus ofclaim 21, wherein a signature is respectively created for each of said plurality of high resolution patches included in said second patch library by generating a feature vector there for that includes an average color for a respective one of said plurality of high resolution patches.

28. The apparatus ofclaim 27, wherein said average color included in said feature vector for said respective one of said plurality of high resolution patches is further of surrounding pixels with respect to said respective one of said plurality of high resolution patches.

29. The apparatus ofclaim 21, wherein only patches preceding a co-located patch with respect to said corresponding one of said plurality of overlapping blocks are used for said searching.

30. The apparatus ofclaim 21, wherein said metadata comprises a patch index for said best matching patch when said difference between said corresponding one of said plurality of overlapping blocks and said best matching patch is less than said threshold difference, said metadata further comprising a block identifier for said pruned block.

Images