CN105574532B - A method and system for region feature description based on binary segmentation tree - Google Patents

Abstract

The present invention is suitable for image characteristics extraction, provides a kind of provincial characteristics based on binary segmentation tree and describes method, step includes: A, and original image is divided into several cut zone；Highest two adjacent area of similarity is synthesized a superzone domain by B, the similarity calculated between adjacent cut zone；C repeats B until being merged into complete original image；D constructs binary segmentation tree according to cut zone and superzone domain；E extracts the feature in the corresponding superzone domain in target area and target area in binary segmentation tree respectively and constructs provincial characteristics accordingly, then described according to the feature that provincial characteristics carries out target area.The present invention has merged the fundamental region feature in target area, superzone domain at the same level and neighbouring higher level region and superzone domain to describe the feature of target area, both local features are contained, global context information is also contained, so the advantage that can have context aware power, high resolution, robustness strong.

Description

A kind of provincial characteristics based on binary segmentation tree describes method and system

Technical field

The invention belongs to computer vision and field of image processing more particularly to a kind of region based on binary segmentation tree are specialSign description method and system.

Background technique

Feature description is a basic module in computer vision and image procossing, in the detection of image well-marked target, meshIn the application such as mark segmentation, target identification and semantic image retrieval, the robustness of feature and rich largely determineIts performance.The feature extracting method of current most of uses is: will be former by the method that image segmentation algorithm or pixel clusterThen beginning image segmentation extracts characteristics of image according to these image subblocks or image-region at image subblock or image-region.

Although existing feature extracting method has high efficiency, natural image often all contains the back of many complexityScape, present state-of-the-art image partition method completely can't also isolate the specific target of boundary definition from backgroundCome, when handling the image of background complexity, does not have enough diversity factoies using the extracted feature of these methods, adopt at presentFacilitate its task with target is formed by less region by adjusting partitioning parameters, but for many images, adjustment pointCutting parameter goes the quantity for reducing region to may result in the less divided phenomenon that image target area merges with background area, to leadThe testing result of mistake is caused, while the prior art, when carrying out feature extraction, feature robustness is poor, information content is not abundant enough.

Summary of the invention

Technical problem to be solved by the present invention lies in provide a kind of provincial characteristics based on binary segmentation tree to describe methodAnd system, it is intended to solve the prior art when carrying out feature extraction, the problem that feature robustness is poor, information content is not abundant enough.

The invention is realized in this way a kind of provincial characteristics based on binary segmentation tree describes method, step includes:

Original image is divided into several cut zone by step A；

Step B calculates the similarity between the adjacent cut zone, and highest two adjacent area of similarity is synthesizedOne superzone domain；

Step C repeats step B until obtained all adjacent areas are merged into complete original image；

Step D constructs binary segmentation tree according to the cut zone and the superzone domain；

Step E extracts the corresponding superzone domain in target area and the target area in the binary segmentation tree respectivelyFeature simultaneously constructs provincial characteristics accordingly, is then described according to the feature that the provincial characteristics carries out target area.

Further, in step, the original image is split using image segmentation algorithm.

Further, in stepb, the similarity between the adjacent cut zone is calculated, similarity is highestTwo cut zone merge to obtain superzone domain, specifically include:

With R_jIndicate a certain cut zone, N_jIndicate cut zone R_jAdjacent area set, S (R_n,R_j) indicate to divideCut region R_jWith its a certain adjacent area R_nSimilarity, R_sIndicate cut zone R_jSimilarity between region adjacent theretoMaximum value, then the region merging technique criterion of two cut zone are as follows:

R_s=arg max S (R_n,R_j), R_n∈N_j。

It further, include root node, leaf node and non-leaf nodes in the binary segmentation tree, wherein rootNode indicates that the entire original image merged, leaf node indicate a cut zone, and non-leaf node indicates a superzoneDomain.

Further, the step E is specifically included:

Step E1 extracts the fundamental region feature of target area and the corresponding superzone domain in the target area respectively；

Step E2, with provincial characteristics described in the fundamental region feature construction that is extracted in step E1, according to the provincial characteristicsFeature description is carried out to the target area.

Further, the step E2 is specifically included:

Step E211, calculate the target area in the binary segmentation tree fundamental region feature and the target areaThe fundamental region feature in the superzone domain of corresponding β superior node；

Step E212, according to fundamental region feature construction first area feature；

With R_kIndicate the target area, k indicates that k-th of node, β indicate target area R_kThe grade of corresponding superior nodeNot Shuo, h_kIndicate target area R_kCorresponding fundamental region feature,Indicate first area feature；

Step E213, if the target area R_kThe number of levels for reaching root node in the binary segmentation tree is greater than the meshThe number of levels β of the corresponding superior node in region is marked, then the first area featureAre as follows:

Step E214, if the target area R_kThe number of levels for reaching root node in the binary segmentation tree is less than the meshThe number of levels β, n of the corresponding superior node in mark region indicate that n-1 to the number of levels of the root node, is then replicated in the target areaThe fundamental region feature of node makes the series of superior node be β, i.e., the described first area featureAre as follows:

Further, the step E2 is specifically included:

Step E221, calculate the target area of the binary segmentation tree fundamental region feature and the target area pairThe base region of the corresponding level region of fundamental region feature and the target area in the superzone domain for the β superior node answeredThe fundamental region feature of the level region in characteristic of field and the superzone domain；

Step E222, all fundamental region feature construction second area features calculated according to step E221；

With R_mIndicate the target area, m indicates that m-th of node, β indicate target area R_mThe grade of corresponding superior nodeNot Shuo, h_mIndicate target area R_mCorresponding fundamental region feature,Indicate the target area and superzone domainThe first area feature of building, slibing (h_m) indicate the target area level region fundamental region feature,Indicate target area R_mThe fundamental region feature in corresponding β -1 layers of higher level superzone domain,Indicate second area feature；

Step E223, if the target area R_mThe number of levels for reaching root node in the binary segmentation tree is greater than the meshMark region R_mThe number of levels β of corresponding superior node, then the second area featureAre as follows: then:

Step E224, if the target area R_mThe number of levels for reaching root node in the binary segmentation tree is less than the meshMark region R_mThe number of levels β of corresponding superior node, m indicate the target area to the number of levels of the root node, then it is describedSecond area featureAre as follows:

The present invention also provides a kind of provincial characteristics based on binary segmentation tree to describe system, comprising:

Cutting unit, for original image to be divided into several cut zone；

Combining unit is adjacent by similarity highest two for calculating the similarity between the adjacent cut zoneRegion synthesizes a superzone domain and continues to merge then using the superzone domain of synthesis as the adjacent area of the cut zone, untilTo all adjacent areas be merged into complete original image；

Construction unit, for constructing binary segmentation tree according to the cut zone and the superzone domain；

Unit is described, for extracting target area in the binary segmentation tree respectively and the target area is corresponding surpassesThe feature in region simultaneously constructs provincial characteristics accordingly, is then described according to the feature that the provincial characteristics carries out target area.

Further, the cutting unit is split the original image using image segmentation algorithm；

It include root node, leaf node and non-leaf nodes in the binary segmentation tree, wherein root node indicates to closeAnd complete entire original image, leaf node indicate a cut zone, non-leaf node indicates superzone domain；

The combining unit, for calculating the similarity between the adjacent cut zone, by similarity highest twoA cut zone merges to obtain superzone domain, specifically includes:

With R_jIndicate a certain cut zone, N_jIndicate cut zone R_jAdjacent area set, S (R_n,R_j) indicate to divideCut region R_jWith its a certain adjacent area R_nSimilarity, R_sIndicate cut zone R_jSimilarity between region adjacent theretoMaximum value, then the region merging technique criterion of two cut zone are as follows:

R_s=arg max S (R_n,R_j), R_n∈N_j。

Further, description unit is specifically used for:

Firstly, extracting the fundamental region feature of target area and the corresponding superzone domain in the target area respectively；

Secondly, provincial characteristics described in fundamental region feature construction with extraction；

Finally, carrying out feature description to target area according to the provincial characteristics.

Compared with prior art, the present invention beneficial effect is: the present invention merged target area, superzone domain at the same level andThe feature that target area is described adjacent to the fundamental region feature in higher level region and superzone domain, both contains local features,Global context information is also contained, so the advantage that can have context aware power, high resolution, robustness strong.Thus overcomeThe disadvantage that feature robustness is poor in prior art, information content is not abundant enough.

Detailed description of the invention

Fig. 1 is the flow chart that a kind of provincial characteristics based on binary segmentation tree provided in an embodiment of the present invention describes method.

Fig. 2 is provincial characteristics structural schematic diagram of the building based on binary segmentation tree provided in an embodiment of the present invention.

Fig. 3 is the detailed process that a kind of provincial characteristics based on binary segmentation tree provided in an embodiment of the present invention describes methodFigure.

Fig. 4 is provided in an embodiment of the present invention using without feature structure notable figure generated.

Fig. 5 is the structural representation that a kind of provincial characteristics based on binary segmentation tree provided in an embodiment of the present invention describes systemFigure.

Specific embodiment

In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, rightThe present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, andIt is not used in the restriction present invention.

In the present invention, the method and system that the provincial characteristics based on binary segmentation tree describes again are proposed, are containedThe part of cut zone and global context information, with context aware power, high resolution and the strong advantage of robustness.

The invention is realized in this way a kind of provincial characteristics based on binary segmentation tree as shown in Figure 1 describes method, walkSuddenly include:

Original image is divided into several cut zone by S1.In this step, Mean-shift, Graph are utilizedOne such partitioning algorithm of any image segmentation algorithm such as based, gPb is split the original image.

S2 calculates the similarity between the adjacent cut zone, by the highest two adjacent areas synthesis one of similaritySuperzone domain.

With R_jIndicate a certain cut zone, N_jIndicate cut zone R_jAdjacent area set, S (R_n,R_j) indicate to divideCut region R_jWith its a certain adjacent area R_nSimilarity, R_sIndicate cut zone R_jSimilarity between region adjacent theretoMaximum value, then the region merging technique criterion of two cut zone are as follows:

R_s=arg max S (R_n,R_j), R_n∈N_j。

S3 repeats step S2 until obtained all adjacent areas are merged into complete original image in this stepIn, the superzone domain that step S2 is synthesized repeats step S2 until being merged into as the adjacent area of the cut zoneWhole original image；

S4 constructs binary segmentation tree according to the cut zone and the superzone domain.In this step, in the binary pointCutting in tree includes root node, leaf node and non-leaf nodes, wherein and root node indicates the entire original image merged,Leaf node indicates a cut zone, and non-leaf node indicates superzone domain.

S5 extracts the feature of target area and the corresponding superzone domain in the target area respectively, constructs provincial characteristics, thenThe feature description of target area is carried out according to the provincial characteristics.In this step, after building binary segmentation tree according to S2,Before carrying out region description, first selected target region, which can be any node on binary segmentation tree.

As shown in Fig. 2, being the provincial characteristics structural schematic diagram based on binary segmentation tree, illustrate how that combination is basic in figureProvincial characteristics constructs target area R_k、R_mAnd R_nCharacteristic area descriptor.In Fig. 2, h_iIndicate target area R_iBase regionCharacteristic of field.

When constructing the frame of binary segmentation tree, it is necessary first to carry out region segmentation in original image, be built in region segmentationA binary segmentation tree is stood, to extract layered image region and systematically analyze their proximity relations.In practical applicationIn, to the method for region segmentation, there is no limit for example can use Mean-shift, Graph based, gPb etc..Region merging techniqueCriterion is determined by the similarity of adjacent area:

R_s=arg max S (R_n,R_j),R_n∈N_j

N_jRepresent adjacent target region R_jSet, S (R_n,R_j) represent adjacent area R_nAnd R_jSimilarity.Pass through mergingThe region of generation is known as superzone domain, and in binary segmentation tree, root node indicates the entire original image merged, leaf node generationOne cut zone of table, non-leaf nodes represent superzone domain.

On the basis of the binary segmentation tree that such as Fig. 2 is constructed, by following observation, we have proposed an advanced spiesLevy structural framing.

Firstly, the image-region (segment) in cut tree leaf node is very little, people can not be by individually observingAnd effectively recognize them, therefore go to extract feature to be not have for conspicuousness target detection only by small regionEnough diversity factoies；Secondly, complete target (little girl in such as Fig. 2) is in some cut zone or superzone in cut treeIt fails to separate from background completely in domain.But when from we are from leaf node up, the major part of targetIt can become more and more clear, such as arm, skirt and the head of little girl.This illustrates that Two-component Multi-layer expression can capture targetGlobal scene.Based on described above, the present embodiment propose two methods go to calculate leaf node region based on binary segmentationThe provincial characteristics (being named as hierarchy-associated rich features, HARF) of tree.

HARF in structure figures 2₁Specific steps include:

A1 calculates the fundamental region feature and the corresponding β in the target area of the target area in the binary segmentation treeThe fundamental region feature in the superzone domain of a superior node；

A2, according to fundamental region feature construction first area feature；

With R_kIndicate the target area, k indicates that k-th of node, β indicate target area R_kThe grade of corresponding superior nodeNot Shuo, h_kIndicate target area R_kCorresponding fundamental region feature,Indicate first area feature；

A3, if the target area R_kThe number of levels for reaching root node in the binary segmentation tree is greater than the target areaThe number of levels β of corresponding superior node, the then characteristic areaAre as follows:

A4, if the target area R_kThe number of levels for reaching root node in the binary segmentation tree is less than the target areaThe number of levels β, n of corresponding superior node indicate that n-1 node to the number of levels of the root node, is then replicated in the target areaFundamental region feature makes the series of superior node be β, i.e., the described characteristic areaAre as follows:

During constructing HARF1, firstly, calculating the target area R in cut tree_kWith corresponding β superior nodeSuperzone domain fundamental region featureThen the fundamental region feature extracted is piled into an eigenmatrix.ThisThe feature extracted is denoted asWherein, R represents target area, β represent for calculate HARF in dividing layerSuperzone domain number of levels.Thus, it is supposed that the fundamental region for representing each region or superzone domain is characterized in a d dimensional feature squareBattle array, then generateThe dimension of feature is d+d* β.Calculate featureMethod indicate are as follows:

HARF in structure figures 2₂Specific steps include:

B1, fundamental region feature and the target area for calculating the target area of the binary segmentation tree are β correspondingThe fundamental region feature of the corresponding level region of fundamental region feature and the target area in the superzone domain of superior node,And the fundamental region feature of the level region in the superzone domain；

B2, all fundamental region feature construction second area features calculated according to step b1；

With R_mIndicate the target area, m indicates that m-th of node, β indicate target area R_mThe grade of corresponding superior nodeNot Shuo, h_mIndicate target area R_mCorresponding fundamental region feature,Indicate the target area and superzone domainThe first area feature of building, slibing (h_m) indicate the target area level region fundamental region feature,Indicate target area R_mThe fundamental region feature in corresponding β -1 layers of higher level superzone domain,Indicate second area feature；

B3, if the target area R_mThe number of levels for reaching root node in the binary segmentation tree is greater than the target areaR_mThe number of levels β of corresponding superior node, the then characteristic areaAre as follows: then:

B4, if the target area R_mThe number of levels for reaching root node in the binary segmentation tree is less than the target areaR_mThe number of levels β of corresponding superior node, m indicate the target area to the number of levels of the root node, then the characteristic areaDomainAre as follows:

During constructing HARF2, a target area R is given_m, comprehensively consider the superior node in binary segmentation treeSuperzone domainAnd their corresponding level regionsIt is proposed that one can obtainThe method of richer HARF feature, is denoted as

Be byFeature adds the level region silbing (h of target area_m) andThe level region in higher level superzone domainThe new matrix that piles of essential characteristic, expression formula can determineJustice is as follows, and this feature matrix dimension is d+2d* β.

It is to be noted that the complete image in root node had both contained conspicuousness target or had included background area, becauseAnd it is not used to calculate HARF value, in addition, not all region in leaf node reaches the number of levels of root node allIt is equal, thus removes calculating HARF without enough series there are some regions, it in this case, can be by replicating mostThe fundamental region feature in high-rise superzone domain and the HARF feature that target area is generated in conjunction with other features of lower level are such as schemedIn 2WithIn conclusion generating the flow chart of the provincial characteristics based on binary segmentation treeIt can be shown in Fig. 3

It is compared with existing technology, HARF provided by the invention is because merged target area, superzone domain at the same level and neighbourThe fundamental region feature in nearly higher level region and superzone domain describes the feature of target area, both contains local features,Global context information is contained, so the advantage that can have context aware power, high resolution, robustness strong.Thus overcomeHave the shortcomings that feature robustness is poor in technology, information content is not abundant enough.

We are applied in conspicuousness target detection by the experimental verification advantage of HARF, traditional special with usingSign (include color, texture and region property) and convolutional neural networks (CNN) Characteristic Contrast, in MSRA-B, PASCAL-1500 withIt is tested on SOD data set, using F-score (β²=0.3) result is evaluated with mean absolute error (MAE), Fig. 4Different characteristic structure notable figure generated is used to be different, can significantly see the advantage of HARF.Fig. 4 uses different characteristic knotNotable figure caused by structure, wherein (a) indicates original image, (b) indicates traditional characteristic, (c) indicates based on traditional characteristicHARF₁, (d) indicate the HARF based on traditional characteristic₂, (e) indicate traditional characteristic and CNN feature, (f) indicate to be based on traditional characteristicWith the HARF of CNN feature₁, (g) indicate the HARF based on traditional characteristic and CNN feature₂。

In the following table, we have obtained the F for generating notable figure using different characteristic structure on different data sets_βScoreMAE value, as a result in as can be seen that be added HARF feature structure frame after three data sets F_βScore is obviously improved (higher tableShow that performance is better), and mean absolute error has apparent reduction (lower expression performance is better).We pass through theory analysisThe provincial characteristics based on binary segmentation tree proposed all verified with experiment describes the superiority of method (HARF).Following table indicatesThe F of different characteristic structure is used on three data sets_βScore and MAE value.

The present invention also provides a kind of provincial characteristics based on binary segmentation tree as shown in Figure 5 to describe system, comprising:

Cutting unit 1, for original image to be divided into several cut zone；

Combining unit 2 is adjacent by similarity highest two for calculating the similarity between the adjacent cut zoneRegion synthesizes a superzone domain and continues to merge then using the superzone domain of synthesis as the adjacent area of the cut zone, untilTo all adjacent areas be merged into complete original image；

Construction unit 3, for constructing binary segmentation tree according to the cut zone and the superzone domain；

Unit 4 is described, it is corresponding for extracting target area in the binary segmentation tree and the target area respectivelyThe feature in superzone domain simultaneously constructs provincial characteristics accordingly, is then described according to the feature that the provincial characteristics carries out target area.

Further, cutting unit 1 divides image segmentation algorithm to institute using Mean-shift, Graph based, gPb etc.Original image is stated to be split；

It include root node, leaf node and non-leaf nodes in the binary segmentation tree, wherein root node indicates to closeAnd complete entire original image, leaf node indicate a cut zone, non-leaf node indicates superzone domain；

Combining unit 2, for calculating the similarity between the adjacent cut zone, by highest two points of similarityIt cuts region merging technique and obtains superzone domain, specifically include:

With R_jIndicate a certain cut zone, N_jIndicate cut zone R_jAdjacent area set, S (R_n,R_j) indicate to divideCut region R_jWith its a certain adjacent area R_nSimilarity, R_sIndicate cut zone R_jSimilarity between region adjacent theretoMaximum value, then the region merging technique criterion of two cut zone are as follows:

R_s=arg max S (R_n,R_j), R_n∈N_j。

Further, description unit 4 is specifically used for:

Firstly, extracting the fundamental region feature of target area and the corresponding superzone domain in the target area respectively；

Secondly, provincial characteristics described in fundamental region feature construction with extraction；

Finally, carrying out feature description to target area according to the provincial characteristics.

The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the inventionMade any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.

Claims (10)

Translated fromChinese

1.一种基于二元分割树的区域特征描述方法，其特征在于，所述区域特征描述方法的步骤包括：1. a regional feature description method based on binary segmentation tree, is characterized in that, the step of described regional feature description method comprises:步骤A，将原始图像分割成若干分割区域；Step A, the original image is divided into several divided regions;步骤B，计算相邻的所述分割区域之间的相似度，将相似度最高的两相邻区域合成一超区域；Step B, calculating the similarity between the adjacent segmented regions, and combining the two adjacent regions with the highest similarity into a super-region;步骤C，重复执行步骤B直至得到的所有相邻区域合并成完整的原始图像；Step C, repeat step B until all the obtained adjacent areas are merged into a complete original image;步骤D，根据所述分割区域和所述超区域构建二元分割树；Step D, constructs a binary segmentation tree according to the segmentation region and the super region;步骤E，分别提取所述二元分割树中的目标区域和所述目标区域对应的超区域的基本区域特征并据此构建第一区域特征和第二区域特征，然后根据所述第一区域特征和第二区域特征进行目标区域的特征描述；Step E, respectively extracting the target area in the binary segmentation tree and the basic area feature of the super area corresponding to the target area and constructing the first area feature and the second area feature accordingly, and then according to the first area feature Perform feature description of the target area with the second area feature;所述第一区域特征是根据所述目标区域的基本区域特征及所述超区域的基本区域特征构建的，所述第二区域特征是根据所述目标区域的基本区域特征、所述超区域的基本区域特征，所述目标区域对应的同级区域的基本区域特征及所述超区域的同级区域的基本区域特征构建的。The first region feature is constructed according to the basic region feature of the target region and the basic region feature of the super region, and the second region feature is constructed according to the basic region feature of the target region and the super region feature. The basic region feature is constructed from the basic region feature of the same level region corresponding to the target region and the basic region feature of the same level region of the super region.2.如权利要求1所述的区域特征描述方法，其特征在于，在步骤A中，利用图像分割算法对所述原始图像进行分割。2 . The area feature description method according to claim 1 , wherein, in step A, the original image is segmented by using an image segmentation algorithm. 3 .3.如权利要求1所述的区域特征描述方法，其特征在于，在步骤B中，计算相邻的所述分割区域之间的相似度，将相似度最高的两个分割区域合并得到超区域，具体包括：3. regional feature description method as claimed in claim 1, is characterized in that, in step B, calculates the similarity between adjacent described segmentation regions, merges two segmentation regions with the highest similarity to obtain super region , including:以R_j表示某一分割区域，N_j表示该分割区域R_j的相邻区域的集合，S(R_n,R_j)表示分割区域R_j和其某一相邻区域R_n的相似度，R_s表示分割区域R_j与其相邻区域之间的相似度的最大值，则两个分割区域的区域合并准则为：Let R_j represent a segmented region, N_j represent the set of adjacent regions of the segmented region R_j , and S(R_n , R_j ) represent the similarity between the segmented region R_j and one of its adjacent regions R_n , R_s represents the maximum similarity between the segmented region R_j and its adjacent regions, then the region merging criterion of the two segmented regions is:R_s＝arg max S(R_n,R_j)，R_n∈N_j。R_s =arg max S(R_n ,R_j ), R_n ∈ N_j .4.如权利要求1所述的区域特征描述方法，其特征在于，在所述二元分割树中包含有根节点、叶子节点和非叶子节点，其中，根节点表示合并完的整个原始图像，叶子节点表示一个分割区域，非叶子结点表示一个超区域。4. The area feature description method according to claim 1, wherein the binary segmentation tree includes a root node, a leaf node and a non-leaf node, wherein the root node represents the entire original image that has been merged, Leaf nodes represent a segmented region, and non-leaf nodes represent a superregion.5.如权利要求4所述的区域特征描述方法，其特征在于，所述步骤E具体包括：5. The area feature description method according to claim 4, wherein the step E specifically comprises:步骤E1，分别提取目标区域和所述目标区域对应的超区域的基本区域特征；Step E1, extract the target area and the basic area features of the super area corresponding to the target area respectively;步骤E2，以步骤E1中提取的基本区域特征构建所述区域特征，根据所述区域特征对所述目标区域进行特征描述。Step E2, constructing the region feature based on the basic region feature extracted in step E1, and characterizing the target region according to the region feature.6.如权利要求5所述的区域特征描述方法，其特征在于，所述步骤E2具体包括：6. The area feature description method according to claim 5, wherein the step E2 specifically comprises:步骤E211，计算所述二元分割树中的目标区域的基本区域特征，及所述目标区域对应的β个上级节点的超区域的基本区域特征；Step E211, calculating the basic region features of the target region in the binary segmentation tree, and the basic region features of the super regions of the β superior nodes corresponding to the target region;步骤E212，根据所述基本区域特征构建第一区域特征；Step E212, constructing a first region feature according to the basic region feature;以R_k表示所述目标区域，k表示第k个节点，β表示目标区域R_k对应的上级节点的级别数，h_k表示目标区域R_k对应的基本区域特征，表示第一区域特征；Let R_k represent the target area, k represents the kth node, β represents the level number of the superior node corresponding to the target area R_k , h_k represents the basic area feature corresponding to the target area R_k , Represents the first region feature;步骤E213，若所述目标区域R_k到达所述二元分割树中根节点的级别数大于所述目标区域对应的上级节点的级别数β，则所述第一区域特征为：Step_E213 , if the number of levels at which the target area Rk reaches the root node in the binary segmentation tree is greater than the number of levels β of the parent node corresponding to the target area, then the first area features for:步骤E214，若所述目标区域R_k到达所述二元分割树中根节点的级别数小于所述目标区域对应的上级节点的级别数β，n表示所述目标区域到所述根节点的级别数，则复制n-1节点的基本区域特征使得上级节点的级数为β，即所述第一区域特征为：Step E214, if the level number β of the target area R_k reaching the root node in the binary segmentation tree is less than the level number β of the parent node corresponding to the target area, n represents the level number β from the target area to the root node , then copy the basic regional features of n-1 nodes so that the series of the superior node is β, that is, the first regional feature for:7.如权利要求5所述的区域特征描述方法，其特征在于，所述步骤E2具体包括：7. The area feature description method according to claim 5, wherein the step E2 specifically comprises:步骤E221，计算所述二元分割树的目标区域的基本区域特征，及所述目标区域对应的β个上级节点的超区域的基本区域特征，以及所述目标区域对应的同级区域的基本区域特征，以及所述超区域的同级区域的基本区域特征；Step E221: Calculate the basic region feature of the target region of the binary segmentation tree, the basic region feature of the super region of the β superior nodes corresponding to the target region, and the basic region of the same level region corresponding to the target region features, and the basic region features of sibling regions of the super region;步骤E222，根据步骤E221计算的所有基本区域特征构建第二区域特征；Step E222, construct the second regional feature according to all the basic regional features calculated in step E221;以R_m表示所述目标区域，m表示第m个节点，β表示目标区域R_m对应的上级节点的级别数，h_m表示目标区域R_m对应的基本区域特征，表示所述目标区域和超区域构建的第一区域特征，slibing(h_m)表示所述目标区域的同级区域的基本区域特征，表示目标区域R_m对应β-1层上级超区域的基本区域特征，表示第二区域特征；_Rm represents the target area,_m represents the_mth node, β represents the number of levels of the superior node corresponding to the target area Rm,_hm represents the basic area feature corresponding to the target area Rm, represents the first area feature constructed by the target area and the super area, slibing(h_m ) represents the basic area feature of the same level area of the target area, Represents the basic regional features of the target region_Rm corresponding to the upper-level superregion of the β-1 layer, Represents the second area feature;步骤E223，若所述目标区域R_m到达所述二元分割树中根节点的级别数大于所述目标区域R_m对应的上级节点的级别数β，则所述第二区域特征为：则：Step E223, if the number of levels at which the target region_Rm reaches the root node in the binary segmentation tree is greater than the level number β of the parent node corresponding to the target region_Rm , then the second region features for: then:步骤E224，若所述目标区域R_m到达所述二元分割树中根节点的级别数小于所述目标区域R_m对应的上级节点的级别数β，m表示所述目标区域到所述根节点的级别数，则所述第二区域特征为：Step E224, if the level number β of the target area R_m reaching the root node in the binary segmentation tree is less than the level number β of the parent node corresponding to the target area R_m , m represents the distance from the target area to the root node. the number of levels, the second area features for:8.一种基于二元分割树的区域特征描述系统，其特征在于，所述区域特征描述系统包括：8. A region feature description system based on a binary segmentation tree, wherein the region feature description system comprises:分割单元，用于将原始图像分割成若干分割区域；a segmentation unit, which is used to segment the original image into several segmented regions;合并单元，用于计算相邻的所述分割区域之间的相似度，将相似度最高的两相邻区域合成一超区域，然后将合成的超区域作为所述分割区域的相邻区域，继续合并，直至得到的所有相邻区域合并成完整的原始图像；The merging unit is used to calculate the similarity between the adjacent segmented areas, combine the two adjacent areas with the highest similarity into a super area, and then use the synthesized super area as the adjacent area of the segmented area, and continue Merge until all adjacent regions obtained are merged into a complete original image;构建单元，用于根据所述分割区域和所述超区域构建二元分割树；a construction unit for constructing a binary segmentation tree according to the segmentation region and the super region;描述单元，用于分别提取所述二元分割树中的目标区域和所述目标区域对应的超区域的基本区域特征并据此构建第一区域特征和第二区域特征，然后根据所述第一区域特征和第二区域特征进行目标区域的特征描述；The description unit is used for extracting the target area in the binary segmentation tree and the basic area feature of the super area corresponding to the target area respectively, and constructing the first area feature and the second area feature accordingly, and then according to the first area feature and the second area feature. The region feature and the second region feature are used to describe the feature of the target region;所述第一区域特征是根据所述目标区域的基本区域特征及所述超区域的基本区域特征构建的，所述第二区域特征是根据所述目标区域的基本区域特征、所述超区域的基本区域特征，所述目标区域对应的同级区域的基本区域特征及所述超区域的同级区域的基本区域特征构建的。The first region feature is constructed according to the basic region feature of the target region and the basic region feature of the super region, and the second region feature is constructed according to the basic region feature of the target region and the super region feature. The basic region feature is constructed from the basic region feature of the same level region corresponding to the target region and the basic region feature of the same level region of the super region.9.如权利要求8所述的区域特征描述系统，其特征在于，所述分割单元利用图像分割算法对所述原始图像进行分割；9. The regional feature description system according to claim 8, wherein the segmentation unit utilizes an image segmentation algorithm to segment the original image;在所述二元分割树中包含有根节点、叶子节点和非叶子节点，其中，根节点表示合并完的整个原始图像，叶子节点表示一个分割区域，非叶子结点表示一个超区域；The binary segmentation tree includes a root node, a leaf node and a non-leaf node, wherein the root node represents the entire original image that has been merged, the leaf node represents a segmentation region, and the non-leaf node represents a super region;所述合并单元，用于计算相邻的所述分割区域之间的相似度，将相似度最高的两个分割区域合并得到超区域，具体包括：The merging unit is used to calculate the similarity between the adjacent segmented regions, and merge the two segmented regions with the highest similarity to obtain a super-region, which specifically includes:以R_j表示某一分割区域，N_j表示该分割区域R_j的相邻区域的集合，S(R_n,R_j)表示分割区域R_j和其某一相邻区域R_n的相似度，R_s表示分割区域R_j与其相邻区域之间的相似度的最大值，则两个分割区域的区域合并准则为：Let R_j represent a segmented region, N_j represent the set of adjacent regions of the segmented region R_j , and S(R_n , R_j ) represent the similarity between the segmented region R_j and one of its adjacent regions R_n , R_s represents the maximum similarity between the segmented region R_j and its adjacent regions, then the region merging criterion of the two segmented regions is:R_s＝arg max S(R_n,R_j)，R_n∈N_j。R_s =arg max S(R_n ,R_j ), R_n ∈ N_j .10.如权利要求8所述的区域特征描述系统，其特征在于，描述单元具体用于：10. The regional feature description system according to claim 8, wherein the description unit is specifically used for:首先，分别提取目标区域和所述目标区域对应的超区域的基本区域特征；First, extract the basic area features of the target area and the super area corresponding to the target area respectively;其次，以提取的基本区域特征构建所述区域特征；Second, construct the region feature with the extracted basic region feature;最后，根据所述区域特征对目标区域进行特征描述。Finally, the target region is characterized according to the region features.