A kind of method for detecting human face and human face detection device, storage mediumTechnical field
The present invention relates to human face detection tech, and in particular to a kind of method for detecting human face and human face detection device, storage are situated betweenMatter.
Background technique
With the development of electronic technology, human face detection and tracing becomes most potential biometric verification of identity means, it is desirable thatAutomatic face recognition system can have certain recognition capability to general pattern, and a series of problems thus faced makes peopleFace is detected initially as an important research topic.Currently, Face datection is a key in Automatic face recognition systemLink, application background is far beyond the scope of face identification system, at content-based retrieval, digital videoReason, video detection etc. have important application value.
Face datection is the fields such as face beauty, face special efficacy, recognition of face, face character analysis, fatigue driving detectionIn necessary pre- pre-treatment step, thus commercial value with higher and application value.But in practical applications, due toBlocked by the expression shape change of face, hair, wear ornaments block, ambient lighting variation, body angle transformation, image-forming condition etc. becauseThe influence of element, so that Face datection is still faced with biggish technological challenge, relevant Face datection algorithm is only obtained furtherWhen ground improves, guarantee could be provided for the practical application effect of Face datection.
Currently, Face datection algorithm can simply be divided into: the Face datection based on the colour of skin, the face based on geometrical characteristicDetection, the Face datection based on statistical learning and the Face datection based on deep learning.Wherein, the face inspection based on deep learningIt surveys over-borrowing and helps deep neural network to reach testing goal, in comparison, accuracy in detection is higher, effect of optimization is obvious,With biggish development prospect.For example, the Face datection based on RCNN series has abandoned the method that sliding window generates candidate region,Using the mode of proposal, although this kind of method can get high detection rate, there are deep neural networks to constitute complicated, inspectionThe slow disadvantage of degree of testing the speed；Face characteristic extracts and classifies often by CNN uniformly come complete in Face datection based on cascade CNN modeAt needing that 6 CNN are arranged in cascade structure, and carry out using 3 CNN face and non-face classification judgement, increaseThe time-consuming of classification judgement, be unfavorable for Face datection fast implements effect.
Summary of the invention
The present invention solves the technical problem of how to improve the Face datection based on deep learning detection speed andDetection accuracy.In order to solve the above technical problems, the application provides a kind of method for detecting human face and its device.
According in a first aspect, providing a kind of method for detecting human face in a kind of embodiment, comprising the following steps:
Image to be detected is obtained in an image sequence；
The processing mode of described image to be detected is selected according to the preceding Face datection result once in described image sequence,When currently once detecting face in described image sequence, then face tracking processing is carried out to described image to be detected, conversely,Recognition of face processing then is carried out to described image to be detected；
The human face region in described image to be detected is exported according to the result of processing.
Face datection result before the basis once in described image sequence selects the processing of described image to be detectedMode, comprising:
Each frame image in described image sequence is successively handled, by the previous frame image of described image to be detectedFace datection result is as previous Face datection as a result, described to be checked according to the previous Face datection result selectionThe processing mode of altimetric image.
It is described using the Face datection result of the previous frame image of described image to be detected as previous Face datection knotFruit, comprising:
Obtain the human face region exported in the previous frame image of described image to be detected；
The human face region that the previous frame image exports is input to a deep neural network for being used for the calculating of face confidence,Obtain face confidence level；
The face confidence level is compared with preset threshold value, when the face confidence level is more than preset threshold valueWhen, then the Face datection result of previous frame image is to detect face, conversely, then the Face datection result of previous frame image is notDetect face.
The human face region by the previous frame image is input to a deep neural network for being used for the calculating of face confidence,Obtain face confidence level, comprising:
The human face region of the previous frame image is zoomed in and out into processing, the image after being scaled；
Image after the scaling is input to a deep neural network for being used for the calculating of face confidence, obtains face confidenceDegree；The deep neural network calculated for face confidence includes one or more bottleneck convolution units, the bottleneck convolutionUnit is used to carry out process of convolution operation to the image of input.
It is described that recognition of face processing is carried out to described image to be detected, comprising:
Described image to be detected is subjected to down-sampled processing, obtains the image of multiple and different sizes；
Each different size of image is input to a lightweight deep neural network for being used for recognition of face, with from describedHuman face region is detected in image to be detected.
It is described that face tracking processing is carried out to described image to be detected, comprising:
The human face region once detected in described image sequence before obtaining；
KCF target is carried out to the preceding human face region once detected in described image sequence in described image to be detectedTracking processing, obtains the human face region in described image to be detected.
It further include frame number judgment step, the frame number is sentenced before carrying out face tracking processing to described image to be detectedDisconnected step includes:
When detecting human face region in described image sequence, each frame image of detection is counted, counting is worked asWhen being as a result more than preset frame number, then ROI region calculating is carried out to described image to be detected, conversely, then to the mapping to be checkedAs carrying out face tracking processing and removing count results to carry out the counting of next round.
It is described that ROI region calculating is carried out to described image to be detected, comprising:
ROI region is carried out to described image to be detected according to the preceding human face region once detected in described image sequenceIt calculates, obtains the discreet area of face in described image to be detected；
The discreet area of face in described image to be detected is input to a lightweight depth nerve for being used for recognition of faceNetwork, to detect human face region from described image to be detected.
The lightweight deep neural network for recognition of face includes:
BP-Net network, for obtaining the candidate region of face in the image of input；
BR-Net network, is trained for the candidate region to the face, and non-face area is removed from candidate regionDomain；
BO-Net network, for being positioned in the candidate region for removing non-face region to face key position, rootHuman face region is obtained according to the positioning result of face key position.
Cascade structure is formed between the BP-Net network, the BR-Net network and the BO-Net network, respectivelyIt include one or more bottleneck convolution units in a network, the bottleneck convolution unit is used to carry out at convolution the image of inputReason operation.
According to second aspect, a kind of embodiment provides a kind of human face detection device, comprising:
Image acquisition unit, for obtaining image to be detected in an image sequence；
Judging unit, for selecting the mapping to be checked according to the preceding once Face datection result in described image sequenceThe processing mode of picture；
Recognition of face processing unit, when for currently once detecting human face region in described image sequence, then to instituteIt states image to be detected and carries out recognition of face processing；
Face tracking processing unit, it is when for currently once not detecting human face region in described image sequence, then rightDescribed image to be detected carries out face tracking processing；
Output unit exports the human face region in described image to be detected according to the result of processing.
According to the third aspect, a kind of computer readable storage medium is provided in a kind of embodiment, which is characterized in that including journeySequence, described program can be executed by processor to realize method as described in relation to the first aspect.
The beneficial effect of the application is:
A kind of method for detecting human face and human face detection device, storage medium according to above-described embodiment.In a first aspect, proposingMethod for detecting human face in joined selection mechanism, through previous testing result from recognition of face processing and face trackingA kind of preferably processing method is selected in reason, conducive to the practical function for enhancing the method for detecting human face；Second aspect, due in peopleThe lightweight deep neural network that recognition of face is introduced in face identification processing procedure passes through BP-Net network therein, BR-Net network and BO-Net network effectively identify and position human face region, conducive to the accuracy for improving Face datection；TheThree aspects, due to introducing KCF target tracking algorism during face tracking, so that the detection process of human face region is quickChange, conducive to the delivery efficiency for improving human face region；Fourth aspect, due to introducing face confidence during human face region exportsDetection method is spent, is conducive to solve drifting problem caused by the face tracking stage, human face region tracing deviation is corrected, to improveThe output prograin of human face region；5th aspect, the application joined on the basis that lightweight deep neural network constructsThe prediction technique of region of interest ROI goes out face in frame image conducive to according to the position prediction of face in previous frame imageRegion that may be present, so that recognition of face is carried out to region that may be present, in this way, can avoid carrying out face to whole imageThe caused situation taken a long time of identification reduces the expense of system conducive to the execution speed for improving recognition of face processing.ThisOutside, the human face detection device that the application proposes has the advantages that structure is simple, algorithmic stability, conducive to and Embedded Hardware Platform phaseIn conjunction with the operation for carrying out Face datection.
Detailed description of the invention
Fig. 1 is a kind of structure chart of the human face detection device of embodiment；
Fig. 2 is a kind of flow chart of the method for detecting human face of embodiment；
Fig. 3 is the flow chart of face confidence level judgement；
Fig. 4 is the flow chart of face identifying processing；
Fig. 5 is the flow chart of face tracking processing；
Fig. 6 is the structure chart of the human face detection device of another embodiment；
Fig. 7 is the flow chart of the method for detecting human face of another embodiment；
Fig. 8 is the structure of the deep neural network for face confidence calculations；
Fig. 9 is the structure of BP-Net network；
Figure 10 is the structure of BR-Net network；
Figure 11 is the structure of BO-Net network；
Figure 12 is the structure of bottleneck convolution.
Specific embodiment
Below by specific embodiment combination attached drawing, invention is further described in detail.Wherein different embodimentsMiddle similar component uses associated similar element numbers.In the following embodiments, many datail descriptions be in order toThe application is better understood.However, those skilled in the art can recognize without lifting an eyebrow, part of featureIt is dispensed, or can be substituted by other elements, material, method in varied situations.In some cases, this ShenPlease it is relevant it is some operation there is no in the description show or describe, this is the core in order to avoid the application by mistakeMore descriptions are flooded, and to those skilled in the art, these relevant operations, which are described in detail, not to be necessary, theyRelevant operation can be completely understood according to the general technology knowledge of description and this field in specification.
It is formed respectively in addition, feature described in this description, operation or feature can combine in any suitable wayKind embodiment.Meanwhile each step in method description or movement can also can be aobvious and easy according to those skilled in the art instituteThe mode carry out sequence exchange or adjustment seen.Therefore, the various sequences in the description and the appended drawings are intended merely to clearly describe a certainA embodiment is not meant to be necessary sequence, and wherein some sequentially must comply with unless otherwise indicated.
It is herein component institute serialization number itself, such as " first ", " second " etc., is only used for distinguishing described object,Without any sequence or art-recognized meanings.And " connection ", " connection " described in the application, unless otherwise instructed, include directly andIt is indirectly connected with (connection).
Embodiment one:
Referring to FIG. 1, the application discloses a kind of human face detection device 1 comprising image acquisition unit 11, judging unit12, recognition of face processing unit 13, face tracking unit 14 and output unit 15 illustrate separately below.
Image acquisition unit 11 in an image sequence for obtaining image to be detected, and in one embodiment, image obtainsUnit 11 obtains a frame image from video flowing, and using the frame image as image to be detected, video flowing here here can beThe video of the monitoring probe shooting of public place can also be the video of the electronic equipments such as mobile phone, camera shooting, and captured viewFrequency includes the video of captured in real-time and the video of passing archive.
Judging unit 12 is used to select the place of image to be detected according to the preceding Face datection result once in image sequenceReason mode.In one embodiment, face identification device 1 successively handles each frame image in image sequence, i.e., according to viewThe time sequencing of image in frequency stream, obtains a frame image every time and (Face datection process can to frame image progress Face datectionWith reference to method for detecting human face hereinafter), using the Face datection result of the previous frame image of image to be detected as previousFace datection is as a result, select the processing mode of image to be detected according to previous Face datection result.It should be noted that working asImage to be detected is first frame image, the error of previous Face datection process or the inspection of previous face in video sequenceSurvey in result when not exporting human face region, then judging unit 12 will judge previous Face datection result be it is no, i.e., do not detectTo face.The specific implementation process of judging unit 12 can refer to method for detecting human face hereinafter.
(judge when recognition of face processing unit 13 is for currently once detecting human face region in described image sequenceThe judging result of unit 12 is yes), then recognition of face processing is carried out to image to be detected.In one embodiment, at recognition of faceIt manages unit 13 and image to be detected is subjected to down-sampled processing, obtain the image of multiple and different sizes；By each different size of figureAs being input to a lightweight deep neural network for recognition of face, to detect human face region from image to be detected.
(i.e. judgement is single when face tracking processing unit 14 in image sequence for not detecting human face region once currentlyThe judging result of member 12 is no), then face tracking processing is carried out to image to be detected.In one embodiment, face tracking is handledThe human face region that unit 14 once detects in image sequence before obtaining；To preceding once in image sequence in image to be detectedIn the human face region that detects carry out KCF target following processing, obtain the human face region in image to be detected.
Result of the output unit 15 for processing exports the human face region in image to be detected.In one embodiment, it exportsHuman face region is carried out rectangle marked by unit 15, and the rectangle marked of human face region is shown in image to be detected.
It will be understood by those of skill in the art that human face detection device 1 has faster place to frame image every in video flowingSpeed (the usually tens frames processing speed per second to several hundred frames per second) is managed, then, output unit 15 can continuously export faceRegion, when user observes video flowing and human face region by display interface, it will be seen that the rectangle marked of human face region is in video flowingThe middle effect for carrying out dynamic mobile.
Correspondingly, referring to FIG. 2, a kind of method for detecting human face is also disclosed in the application comprising step S100-S500, underFace illustrates respectively.
Step S100 obtains image to be detected in an image sequence.In one embodiment, image acquisition unit 11 is oneA frame image is obtained in video flowing, using the frame image as image to be detected.
Step S200 selects the processing side of image to be detected according to the preceding Face datection result once in image sequenceFormula.In one embodiment, judging unit 12 successively handles each frame image in image sequence, by the upper of image to be detectedThe Face datection result of one frame image is as previous Face datection as a result, selecting institute according to previous Face datection resultThe processing mode for stating image to be detected, is shown in Fig. 3, and step S200 may include step S210-S270.
Step S210, judging unit 12 obtain the previous frame image of image to be detected in image sequence.
Step S220, judging unit 12 judges whether output human face region in previous frame image, if so, entering stepS230, conversely, then entering step S270.It should be noted that when image to be detected be video sequence in first frame image,When previous Face datection process malfunctions, judging unit 12 also once will not detect face before judgement.
Step S230, zooms in and out processing for the human face region exported in previous frame image, and the image after being scaled is excellentThe human face region exported in previous frame image is scaled to the image of 12 × 12 pixels by selection of land.
Step S240, the image after scaling, which is input to the deep neural network that one is used for the calculating of face confidence, (can use symbolNumber FCNET is indicated), obtain face confidence level (available symbols C expression).In one embodiment, see what the face confidence of Fig. 8 calculatedDeep neural network, the deep neural network FCNET which calculates include one or more bottleneck convolution units (preferablyGround includes four bottleneck convolution units, respectively two 16 channels, two 24 channels), bottleneck convolution unit master hereIt is used to carry out process of convolution operation to the image of input；Visible Figure 12 of specific structure of each bottleneck convolution unit, wherein BNBelong to the prior art for doing normalized to each neuron for normalized function；RELU is activation primitive, is usedIn the high efficiency for guaranteeing training process, belongs to the prior art, be no longer described in detail here.For the meter for improving face confidence levelThe accuracy of calculation process also joined 3 × 3 × 3 filtering in the present embodiment in the deep neural network that face confidence calculatesThe 2d convolutional coding structure (the 2d convolutional coding structure is mainly used for feature extraction operation) in device and 32 channels and 1 × 1 convolution algorithm listMember.
Face confidence level C and preset threshold value (can be indicated with FT) are compared, as face confidence level C by step S250When greater than preset threshold value FT, then S260 is entered step, otherwise enters step S270.It should be noted that accurately to be sentencedDisconnected, preset threshold value preferably uses 0.93 in the present embodiment.
Step S260, it is believed that the Face datection result of previous frame image is to detect face, i.e. the judgement of judging unit 12 knotFruit is yes.
Step S270, it is believed that the Face datection result of previous frame image is not detect face, i.e. judging unit 12 is sentencedDisconnected structure is no.
Step S300 when currently once detecting human face region in image sequence, then carries out face to image to be detectedIdentifying processing.In one embodiment, image to be detected is carried out down-sampled processing by recognition of face processing unit 13, obtain it is multiple notWith the image of size, each different size of image is input to a lightweight deep neural network for being used for recognition of face, withHuman face region is detected from image to be detected, then, step S300 may include step S310-S330, be respectively described below.
Step S310 after recognition of face processing unit 13 carries out down-sampled processing to image to be detected, forms image goldImage pyramid is preferably divided into three grades by tower, is respectively formed 48 × 48 resolution ratio, 24 × 24 resolution ratio and 12 × 12 pointsThe image of resolution, to obtain multiple images of different sizes.The image of those different resolutions is for adapting to heterogeneous networks knotThe input demand of structure.
Each different size of image is input to one and is used for recognition of face by step S320, recognition of face processing unit 13Lightweight deep neural network (available symbols BFACENET expression).It should be noted that lightweight deep neural networkBFACENET refers to that network forms the less neural network of the number of plies, usually in 10 layers of network below.
In one embodiment, lightweight deep neural network BFACENET includes BP-Net network, BR-Net network and BO-Net network.Wherein, the convolution frame of BP-Net network is as shown in table 1, the visible Fig. 9 of corresponding with table 1 convolutional coding structure.
The convolution frame of 1 BP-Net network of table
In upper table 1, t indicates that expansion multiple, c indicate that output channel number, n indicate unit number, and s indicates span.Each volumeProduct unit preferably uses bottleneck convolutional coding structure, and visible Figure 12 of composition substantially of bottleneck convolutional coding structure.It should be noted that thisIn embodiment, BP-Net network be mainly used for obtaining in 12 × 12 image in different resolution of input face candidate window and return toAmount, to obtain the candidate region of face.
Wherein, the convolution frame of BR-Net network is as shown in table 2, the visible Figure 10 of corresponding with table 2 convolutional coding structure.
The convolution frame of 2 BR-Net network of table
| Input | Convolution operation | " expansion " times t | Port number c | Unit number n | Span s |
| 24x24x3 | Conv2d | - | 8 | 1 | 2 |
| 12x12x8 | Convolution unit | 6 | 16 | 2 | 2 |
| 6x6x16 | Convolution unit | 6 | 24 | 2 | 2 |
| 3x3x24 | 3x3 convolution unit | - | 32 | 2 | 1 |
| 1x1x32 | Conv2d 1x1 | - | 96 | 1 | - |
It should be noted that BR-Net network is mainly used for being trained the candidate region of face, from candidate regionRemove non-face region.In one embodiment, BR-Net network is according to the image of 24 × 24 resolution ratio of input to faceCandidate region be trained, to remove non-face region.
Wherein, the convolution frame of BO-Net network is as shown in table 3, the visible Figure 11 of corresponding with table 3 convolutional coding structure.
The convolution frame of 3 BO-Net network of table
| Input | Convolution operation | " expansion " times t | Port number c | Unit number n | Span s |
| 48x48x3 | Conv2d | - | 8 | 1 | 2 |
| 12x12x8 | Convolution unit | 6 | 16 | 2 | 2 |
| 6x6x16 | Convolution unit | 6 | 24 | 2 | 2 |
| 3x3x24 | 3x3 convolution unit | - | 48 | 2 | 1 |
| 1x1x48 | Conv2d 1x1 | - | 128 | 1 | - |
It should be noted that BO-Net network is mainly used in the candidate region for removing non-face region to face keyPosition is positioned, and obtains human face region according to the positioning result of face key position.In one embodiment, BO-Net netNetwork positions face key position in candidate region according to the image of 12 × 12 resolution ratio of input, from two eye centers,Nose, two corners of the mouths five human body key positions determine face, and obtain human face region.
It will be understood to those skilled in the art that in the lightweight deep neural network BFACENET that the present embodiment usesForm cascade structure between BP-Net network, BR-Net network and BO-Net network, include in each network one orMultiple bottleneck convolution units (available symbols BottleNeck expression), bottleneck convolution unit BottleNeck are used for the figure to inputAs carrying out process of convolution operation, since the structure of bottleneck convolution unit BottleNeck is simple, conducive to constructed network is reducedParameter, from the arithmetic speed for accelerating Face datection.
The objective function of the present embodiment training associated depth neural network model are as follows:
Formula (1-1) is to (1-4), yiRepresent the sample label of face, piFor the probability for obtaining face；Det is face classificationTask (or recurrence task for Face datection), box are that bounding box returns task (or judgement task for face), and mark is indicatedKey point location tasks；∝jIt is returned for face classification, bounding box, three tasks of crucial point location are being presently in stage damageBecome homeless weight (the preferably ∝ in the present embodiment accounted fordet=0.5, ∝box=0.25, ∝mark=0.25)；Whether to be faceInstruction scalar, use 1 indicates face, and use 0 indicates no face；I, j indicates the serial number of current task, and when subscript indicates currentTask category, the stage locating for current task is indicated when subscript；L is loss function；Double absolute value signs indicate secondary normal formIt calculates；In { } indicates set operation.
It should be noted that formula (1-1) to (1-4) can be seen that, the result of upper layer sub-network is by next straton network instituteIt uses, to reach mutual cascade effect between BP-Net network, BR-Net network and BO-Net network.
It should be noted that the face classification in Fig. 9, Figure 10 and Figure 11 is the vector of a 1x1x2, that is to say, that it has1 or 0 result indicates；Bounding box recurrence is the vector of a 1x1x4, the coordinate of main output boundary frame；Crucial point location isThe vector of one 1x1x10, the main coordinate for exporting 5 key points of face.
Step S330.Recognition of face processing unit 13 detects human face region from image to be detected.In one embodiment,Recognition of face processing unit 13 obtains human face region according to the face key position positioned in BO-Net network, using as detectingHuman face region.
Step S400, when currently once detecting human face region in image sequence, to image to be detected carry out face withTrack processing.In one embodiment, the human face region that face tracking processing unit 14 once detects in image sequence before obtaining,KCF target following processing is carried out to the preceding human face region once detected in image sequence in image to be detected, thusHuman face region into image to be detected, then, step S400 may include step S410-S430.
Step S410, the human face region once detected in image sequence before obtaining.In one embodiment, precedingIn the case that one-time detection goes out human face region, face tracking processing unit 14 obtains the previous frame figure of image to be detected in video flowingThe human face region detected as in.
Step S420 carries out KCF target following processing in image to be detected.In one embodiment, face trackingThe human face region detected in previous frame image and image to be detected are input to KCF target following processing by processing unit 14 jointlyAlgorithm among, target following is carried out to the human face region that detects in previous frame image in image to be detected, to obtainHuman face region in image to be detected.
It should be noted that the algorithm of KCF target following processing is more common in field of image processing, it is often used for pairTarget object in image is tracked analysis, and this method is typically all one object detector of training in tracing process,It goes whether detection next frame predicted position is target using object detector, then reuses new testing result and go to update training setAnd then update object detector.In training objective detector, general target area of choosing is positive sample, the peripheral region of targetFor negative sample, a possibility that closer to mesh target area being positive sample, is bigger.This formal for using KCF target following Processing AlgorithmKind of characteristic, the present embodiment realize according to the human face region in previous frame image to the human face region in image to be detected carry out withThe purpose of track.Since KCF target following Processing Algorithm is the prior art, therefore is no longer described in detail here.
Step S430 obtains the human face region in image to be detected.In one embodiment, human face target tracking is handledUnit 14 identifies the human face region obtained according to KCF target following Processing Algorithm, to obtain in image to be detectedHuman face region.
It should be noted that KCF target following Processing Algorithm during track human faces region, can be produced inevitablyRaw drift may cause the accuracy decline of target following, to avoid this situation, in the next frame image to image to be detectedBefore carrying out human face target tracking processing, the face confidence in step S200 is also carried out to the human face region in image to be detectedDegree calculates, and drifting problem brought by KCF target tracking algorism can be effectively prevented by face confidence level, to be stepJudging result in S200 provides accurate judgment basis.
Step S500 exports the human face region in described image to be detected according to the result of processing.In one embodiment, seeFig. 1, output unit 15 is by people in the human face region that recognition of face processing unit 13 detects in step S330 or step S430The human face region that face tracking treatment unit 14 detects carries out rectangle marked, continuously exports each frame image in image sequence,If detecting human face region in the current frame image of output, the rectangle mark of human face region on current frame image and image is exportedNote only exports current frame image if not detecting human face region in the current frame image of output.
Embodiment two:
Referring to FIG. 6, the human face detection device 2 of another embodiment is also disclosed in the application comprising people in embodiment oneFace detection device 1 further includes frame number judging unit 16 and ROI region computing unit 17, illustrates separately below.
The frame number judging unit 16 is located between detection judging unit 12 and face tracking processing unit 14, for schemingWhen as detecting human face region in sequence, each frame image of detection is counted, when count results are more than preset frameWhen number (preset frame number available symbols T is indicated, preferably using the numerical value within the scope of 48-128), then to described image to be detectedROI region calculating is carried out, conversely, then carrying out face tracking processing to image to be detected and removing count results to carry out next roundCounting.
ROI region computing unit 17 is connect with frame number judging unit 16 and recognition of face processing unit 13, in frame numberWhen judging unit 16 judges that calculated result is more than preset frame number, according to the preceding human face region once detected in image sequenceROI region calculating is carried out to image to be detected, obtains the discreet area of face in image to be detected.Then ROI region calculates singleThe discreet area of face in image to be detected is input in recognition of face processing unit 13 by member 17, in the discreet area of faceThe treatment process of the interior lightweight deep neural network for carrying out recognition of face, to detect face area in image to be detectedDomain.
It should be noted that area-of-interest (ROI) is the region selected from image in field of image processing,This region the emphasis of interest as image analysis is drawn a circle to approve into the region to be further processed, when using ROI regionThe target object in image can quickly be obtained, it is possible to reduce the processing time increases processing accuracy.
It should be noted that KCF target following Processing Algorithm employed in face tracking processing unit 14 for a long time,It will lead to drifting problem when continuously handling image, will seriously affect the detection effect of human face region, and frame number judging unit 16It can be carried out continuously in face tracking processing unit 14 after certain face tracking number of processes, ROI is carried out to image to be detectedRegion calculates, to carry out recognition of face processing in ROI region, quickly detects standard conducive within lesser image-regionTrue human face region, and then achieve the purpose that carry out position correction to human face region, it avoids face tracking processing unit 14 and existsFallibility when KCF target following processing and the drifting problem of generation are carried out to human face region.
Referring to FIG. 7, correspondingly, another method for detecting human face is also disclosed in the present embodiment comprising step S100-S600。
Method for detecting human face in the present embodiment two has had more step relative to the method for detecting human face in embodiment oneS600, step S600 may include step S610-S630, illustrate separately below.
Step S610 is located at before step S400, can be described as frame number judgment step.In one embodiment, which judgesStep includes:
It detects to start (i.e. judging result is when being to detection judging unit 12 for the first time) when human face region in image sequence,Each frame image of detection is counted, when count results are more than preset frame number T (preferably using within the scope of 48-128Numerical value), then S620 is entered step, conversely, then entering step S400.
It should be noted that be so that the counting arbitration functions in step S610 persistently carry out, when entering step S400,Frame number judging unit 16 will remove count results to carry out the counting of next round, that is, detect the judging result of judging unit 12 againSecondary is when being, frame number judging unit 16 restarts to count.
Step S620 carries out the area ROI to image to be detected according to the preceding human face region once detected in image sequenceDomain calculates, and obtains the discreet area of face in image to be detected.In one embodiment, calculating process are as follows:
ROIW=T1*FACEW (2-3)
ROIH=T2*FACEH (2-4)
In above formula, ROIXFor the x coordinate of the top left corner pixel point of area-of-interest, ROIYFor the upper left corner of area-of-interestThe y-coordinate of pixel, then, (FACEX, FACEY) be the human face region detected in previous frame image top left corner pixel pointCoordinate.FACEWAnd FACEHThe width and height of the human face region respectively detected in previous frame image, ROIWAnd ROIHTo needThe height and width for the area-of-interest to be calculated；T1, T2 be respectively the threshold value of the customized setting of user (rule of thumb, preferablyT1=2.5, T2=1.6 is arranged in ground).If ROIX、ROIY、FACEWAnd FACEHValue exceed image boundary, then with image boundaryCoordinate value be true value.
The discreet area of face in image to be detected is input to a lightweight depth for being used for recognition of face by step S630Neural network, to detect human face region from image to be detected.In one embodiment, ROI region computing unit 17 will be to be checkedThe discreet area of face is input in face identification unit 13 in altimetric image, and discreet area is input to the lightweight of recognition of faceIn deep neural network BFACENET, to obtain the human face region in image to be detected.The lightweight depth mind of recognition of faceComposition and calculation method through network can refer to the step S300 in embodiment one, and it will not be described here.
It will be understood by those of skill in the art that using the method needs pair of lightweight deep neural network in step S300Entire image to be detected is scanned, and by traversal whole image come locating human face region, therefore will increase many invalid timesBetween lasting.And the characteristics of temporal correlation in image sequence is utilized in step S630, using lightweight deep neural networkMethod only needs to be scanned the image in face discreet area, so that the range of Face datection traversal be schemed by wholeAs narrowing down to the range estimated, the sliding of good-for-nothing's face can be reduced, the interference in non-face region, and then the fortune of boosting algorithm are removedCalculate speed.
It will be understood by those skilled in the art that all or part of function of various methods can pass through in above embodimentThe mode of hardware is realized, can also be realized by way of computer program.When function all or part of in above embodimentWhen being realized by way of computer program, which be can be stored in a computer readable storage medium, and storage medium canTo include: read-only memory, random access memory, disk, CD, hard disk etc., it is above-mentioned to realize which is executed by computerFunction.For example, program is stored in the memory of equipment, when executing program in memory by processor, can be realizedState all or part of function.In addition, when function all or part of in above embodiment is realized by way of computer programWhen, which also can store in storage mediums such as server, another computer, disk, CD, flash disk or mobile hard disksIn, through downloading or copying and saving into the memory of local device, or version updating is carried out to the system of local device, when logicalWhen crossing the program in processor execution memory, all or part of function in above embodiment can be realized.
Use above specific case is illustrated the present invention, is merely used to help understand the present invention, not to limitThe system present invention.For those skilled in the art, according to the thought of the present invention, can also make several simpleIt deduces, deform or replaces.