CN105138975B - A kind of area of skin color of human body dividing method based on degree of depth conviction network - Google Patents
A kind of area of skin color of human body dividing method based on degree of depth conviction network Download PDFInfo
- Publication number
- CN105138975B CN105138975B CN201510497519.XA CN201510497519A CN105138975B CN 105138975 B CN105138975 B CN 105138975B CN 201510497519 A CN201510497519 A CN 201510497519A CN 105138975 B CN105138975 B CN 105138975B
- Authority
- CN
- China
- Prior art keywords
- degree
- depth
- skin
- colour
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 69
- 238000012549 training Methods 0.000 claims abstract description 35
- 239000003086 colorant Substances 0.000 claims abstract description 24
- 238000013480 data collection Methods 0.000 claims abstract description 19
- 238000012360 testing method Methods 0.000 claims abstract description 18
- 239000000284 extract Substances 0.000 claims abstract description 7
- 238000001514 detection method Methods 0.000 claims description 35
- 238000002474 experimental method Methods 0.000 claims description 16
- 230000001537 neural effect Effects 0.000 claims description 13
- 230000000694 effects Effects 0.000 claims description 11
- 238000005457 optimization Methods 0.000 claims description 9
- 238000002790 cross-validation Methods 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000007689 inspection Methods 0.000 claims description 5
- 201000007094 prostatitis Diseases 0.000 claims description 4
- 230000019552 anatomical structure morphogenesis Effects 0.000 claims description 3
- 238000012545 processing Methods 0.000 abstract description 5
- 230000008569 process Effects 0.000 description 13
- 230000011218 segmentation Effects 0.000 description 12
- 230000000877 morphologic effect Effects 0.000 description 7
- 238000013459 approach Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000007500 overflow downdraw method Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000001174 ascending effect Effects 0.000 description 2
- 238000011960 computer-aided design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 230000036541 health Effects 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000003750 conditioning effect Effects 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 239000007933 dermal patch Substances 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000001310 location test Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000012732 spatial analysis Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
- G06V40/162—Detection; Localisation; Normalisation using pixel segmentation or colour matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of area of skin color of human body dividing method based on degree of depth conviction network, belongs to computer vision and image processing field. The method comprises: (1). set up colour of skin blocks of data collection and non-colour of skin blocks of data collection; (2). produce the degree of depth conviction network input feature vector collection in distinct colors space and combination thereof; (3). the degree of depth conviction network of training different scale, and optimize and determine parameters; (4). input test pattern picture, under different scale retrains, it may also be useful to moving window strategy obtains current image block scope to be detected; (5). on the image block that current moving window is determined, extract color space component or its combination of needs, obtain input feature vector vector; (6). input feature vector vector being inputted the degree of depth conviction network trained, obtains the classification results of image block, if classification results is colour of skin block, then block interior pixel point puts 1; If classification results is non-colour of skin block, then block interior pixel point sets to 0.
Description
Technical field
The invention belongs to computer vision and image processing field, it relates to machine learning and human body analyze technology, are specifically related to the area of skin color of human body dividing method based on degree of depth conviction network.
Background technology
Area of skin color of human body is segmented in computer vision process field and has more application, such as can detect human face region fast, in based on the application of gesture interaction, locate staff position, the Attitude estimation and behavioural analysis of health provide the starting position etc. at human body complexion position. In the many application using Face Detection, general requirement skin color detection method can provide under complicated environment Shandong rod, area of skin color of human body comparatively accurately, and do not need result by pixel classifications accurately.
Current skin color segmentation method is mainly based on the method by pixel classifications, such method is taking the pixel of image as process unit, each pixel in image is classified by rule or training sorter by formulating, it is divided into skin pixel point or non-skin pixel point, obtains area of skin color of human body segmentation result finally by morphological operations such as burn into expansions. Obtain discrete skin pixel point based on the method by pixel classifications, not only affect follow-up morphological operation and produce inaccurate result, and this type of method is examined more under complicated environment, directly has influence on the effect of advanced processes step by mistake. Consider that similar pixel point cluster is the feature in region, also have method to propose the skin color segmentation method based on region, but this type of method essence is based on classify of image element result, the effect of skin color segmentation can be had influence on by the shortcoming of pixel detection equally.
" computer aided design (CAD) and graphics journal " " skin color detection method unrelated with color-space choosing " 01 phase in 2013 discloses a kind of skin color detection method based on flexible neural tree, the flexible neural tree of the method stochastic generation different structure, the genetic algorithm guided based on grammer is used to carry out composition optimizes, use particle swarm optimization algorithm parameter to be optimized simultaneously, finally obtain complexion model. In literary composition, the accuracy rate of method and rate of false alarm are better than other main flow skin color detection methods, and the detection of colour of skin details under complicated environment is had good effect.
Within 2014, Springer periodical discloses one section of skin color detection method merged about multiple color spaces, name is called: the multiple color space fusion method (Multiplecolorspacechannelfusionforskindetection) in Face Detection, this article proposes a kind of skin color detection method that different colours spatial component carries out linear fusion, for obtaining the effect to illumination variation Shandong rod. The weight learning process wherein proposed utilizes positive sample training data study color weight model. Experimental result demonstrates color space fusion method can carry out Shandong rod and stable Face Detection.
Within 2014, Springer periodical also discloses one section of skin color detection method merged about multiple color spaces, name is called: the multiple color space fusion method (Systematicskinsegmentation:mergingspatialandnon-spatiald ata) in Face Detection, this literary composition proposes a kind of figure of use and cuts the method that algorithm carries out skin color segmentation, first utilize the local colour of skin information of the face detected as foreground seeds, and the prospect weight for scaling system. If local colour of skin information is unavailable, then select general seed, use the sorter based on decision tree to promote robustness simultaneously, thus obtain the Shandong rod skin color segmentation result being better than off-line sorter. In addition, it may also be useful to the mode of Face datection can effectively utilize scene context information, it may also be useful to weight method for improving merges space and non-space data. Experimental result proves that institute's extracting method is better than additive method and has good robustness.
Within 2014, like that thinking only your periodical discloses a kind of based on space and the skin color detection method differentiating colour of skin representation feature, name is called: Spatial-basedskindetectionusingdiscriminativeskin-presen cefeatures, the method is for based on overlap problem in many color spaces of the limitation of pixels approach and skin pixel and non-skin pixel, it is proposed to utilize textural characteristics and space characteristics to promote the discriminating power of skin color classifier. Propose the spatial analysis differentiating that feature space is used for skin pixel, from skin color probability map, extract textural characteristics. Experimental result shows the skin color detection method that institute's extracting method uses textures and space characteristics due to other.
Within 2013, like that thinking only your periodical also discloses the skin color detection method of a kind of facing area, name is called: Skindetectionbydualmaximizationofdetectorsagreementforvi deomonitoring, this article proposes the skin color detection method of a kind of auto-adaptive parameter, it may also be useful to two kinds of detectors obtain region and discrete pixel detected result respectively. By the operability of the contact lift frame between study detector parameters, two class detectors are combined by morphology reconstruction filtering, have both remained area of skin color and have removed detection mistake region simultaneously. Experiment at Human bodys' response data set proves that institute's extracting method is better than most of methods involving.
In sum, existing area of skin color of human body cutting techniques belongs to the method by pixel classifications mostly, and computer vision field uses the scene of Face Detection to be generally the probable ranges utilizing skin color detection method to provide health colour of skin position, instead of accurate classifying by pixel. Shortcoming by the method for element marking is: accuracy rate is not high on the one hand, examines more so that the result of Face Detection does not reach the requirement of advanced processes by mistake; The detected result of this type of method is discrete pixel one by one on the other hand, in order to obtain area of skin color segmentation result, often need to use morphological operation to carry out subsequent disposal, but carrying out morphological operation on the basis of discrete pixel, result can be had a negative impact by the pixel of inspection by mistake.
Summary of the invention
It is an object of the invention to solve in above-mentioned prior art the difficult problem existed, a kind of area of skin color of human body dividing method based on degree of depth conviction network is provided, using image block as processing unit, the area of skin color detection under complex background is had good robustness by the degree of depth conviction network model of training.
The present invention is achieved by the following technical solutions:
A kind of area of skin color of human body dividing method based on degree of depth conviction network, it is characterised in that: described method comprises:
(1). set up colour of skin blocks of data collection and non-colour of skin blocks of data collection;
(2). produce the degree of depth conviction network input feature vector collection in distinct colors space and combination thereof;
(3). the degree of depth conviction network of training different scale, and optimize and determine parameters;
(4). input test pattern picture, under different scale retrains, it may also be useful to moving window strategy obtains current image block scope to be detected;
(5). on the image block that current moving window is determined, extract color space component or its combination of needs, obtain input feature vector vector;
(6). input feature vector vector being inputted the degree of depth conviction network trained, obtains the classification results of image block, if classification results is colour of skin block, then block interior pixel point puts 1; If classification results is non-colour of skin block, then block interior pixel point sets to 0;
(7). judge whether entire image scanning completes, if not, then according to the sliding window moving step length moving window of setting, then return step (5), if it does, then proceed to step (8);
(8). obtain final Face Detection result, terminate.
As the further restriction to the technical program, described step (1) comprises the following steps:
(11) from the non-colour of skin sample of Compaq data centralization, the image block being of a size of 10x10 and 20x20 size is randomly drawed as non-colour of skin blocks of data collection;
(12) from the colour of skin sample that ECU data is concentrated, the image block being of a size of 10x10 and 20x20 size is randomly drawed as colour of skin blocks of data collection;
(13) data set of 10x10 and 20x20 is divided into 5 equal portions at random, for training the cross validation of degree of depth conviction network.
As the further restriction to the technical program, described step (2) comprises the following steps:
B1: select RGB, HSV, YCbCr and CIELab to be experiment color space, the extraordinary component that each channel components of often kind of color space is concentrated as feature, considers the CbCr color space after removing Y-component simultaneously;
B2: whole combination spaces that color space combines RGB+YCbCr, RGB+HSV, RGB+CIELab, HSV+YCbCr, HSV+CIELab, YCbCr+CIELab, RGB+YCbCr+HSV, RGB+CIELab+HSV and YCbCr+CIELab+HSV and four kinds of spaces are tested respectively, detection depth model combines effect spatially at these;
B3: the color space component extracting all pixels of moving window place image block, is normalized to [0,1] scope, according to different colours space or combination, obtains set of eigenvectors.
As the further restriction to the technical program, described step (3) comprises the following steps:
(31) determine that degree of depth conviction network packet is containing 4 layers, wherein hidden layer three layers, input layer one layer, by three limited Bohr hereby graceful machine form degree of depth conviction network;
(32) degree of depth skin color classifier increases by one layer of full interconnection network on degree of depth conviction network structure basis, for the output of classification results;
(33) sorter hidden layer structure is 100-100-100-2, and training minimum packets is 100, and iteration number of times is 200;
(34) on the data set of 10x10 and 20x20, corresponding degree of depth complexion model DSM is trained respectively10And DSM20��
As the further restriction to the technical program, described step (34) comprises the following steps:
C4.1: obtain present image block feature x=(x0, x1..., xi);
C4.2: taking x=h (0) as degree of depth complexion model input layer, the h1 layer of training model;
C4.3: use h1 as the input data of the second layer;
C4.4: using h1 as learning sample, for training the h2 layer of RBM;
C4.5: iteration C4.3 and C4.4 step, instruct the training of all layers;
C4.6: using counterpropagation network adjusting and optimizing model parameter, wherein each parameter optimisation procedure is as follows:
C4.6.1: optimum color space and combination selection: adopt cross validation mode training and testing degree of depth complexion model respectively on the data set of 10x10 and 20x20, optimal colors space is chosen, simultaneously the degree of depth skin color classifier performance under experiment test different colours spatial array according to accuracy rate, by mistake inspection rate.
C4.6.2: degree of depth conviction network architecture parameters is optimized, i.e. hidden layer neuron quantity optimization: setting neuronal quantity respectively is 100,200 and 500, carries out the experiment of the data set of 20x20 under optimal colors space, finds best neuronal quantity;
C4.6.3: the determination of iteration number of times in optimization: setting iteration number of times respectively is 50,100 and 200, and the data set of 20x20 is tested under optimal colors space, finds best iteration number of times.
As the further restriction to the technical program, described step (4) comprises the following steps:
(41) setting step-length step size respectively is 2 pixels, 5 pixels and 10 pixels;
(42) according to different step-length setting during mobile sliding window, moving window is moved with vertical direction in the horizontal direction, it is determined that current window position.
As the further restriction to the technical program, described step (42) comprises the following steps:
D1: setting moving window size is the tile size that degree of depth complexion model is corresponding;
D2: setting moving window moving step length is 2 pixels, 5 pixels and 10 pixels;
D3: obtain current image length to be detected and wide data;
D4: from the upper left corner, obtains current window scope, and namely when the coordinate row that moves ahead adds length of window: row+patch_size, when prostatitis, coordinate adds window width: col+patch_size;
D5: obtain current image block content:
Patchi=img (row:row+patch_size-1, col:col+patch_size-1 :);
D6: horizontal direction and vertical direction are with step distance moving window, and repeating step D4 and D5, until row and col reaches image maximum length and width.
As the further restriction to the technical program, described step (5) is achieved in that
First determine moving window position, obtain the coordinate of the whole pixels in window area, then extract the channel value of each pixel in distinct colors space or spatial array, i.e. color space component values, morphogenesis characters component of connecting after normalization method.
Compared with prior art, the invention has the beneficial effects as follows:
1., in units of image block instead of taking single pixel as process unit, in testing process, directly consider the spatial context information between pixel;
2., based on the successful Application of degree of depth learning method at computer vision field, in Face Detection is split, the complexion model based on degree of depth conviction network is proposed first;
3. establish a colour of skin block for degree of depth complexion model training and testing and non-colour of skin blocks of data collection;
4. propose the skin color segmentation system of the fusion degree of depth complexion model taking moving window as framework;
5. an extracting method set up image block data collection on performance be better than current main-stream method, by the effect of pixel classifications reaches main stream approach performance.
Accompanying drawing explanation
The step block diagram of Fig. 1 the inventive method.
Fig. 2 degree of depth conviction network structure
Fig. 3 degree of depth complexion model schematic diagram.
Embodiment
Below in conjunction with accompanying drawing, the present invention is described in further detail:
As shown in Figure 1, the inventive method comprises:
1. set up colour of skin blocks of data collection and non-colour of skin blocks of data collection;
2. produce the degree of depth conviction network input feature vector collection in distinct colors space and combination thereof;
3. train the degree of depth conviction network of different scale, and optimize and determine parameters;
4. input test pattern picture, under different scale retrains, it may also be useful to moving window strategy obtains current image block scope to be detected;
5. the image block determined at current moving window extracts color space component or its combination of needs, obtain input feature vector vector (performing step: first determine moving window position, obtain the coordinate of the whole pixels in window area, then the channel value (i.e. color space component values) of each pixel in distinct colors space (or spatial array) is extracted, morphogenesis characters component of connecting after normalization method. );
6. input feature vector vector being inputted the degree of depth conviction network trained, obtain the classification results of image block, if classification results is colour of skin block, then block interior pixel point puts 1; If classification results is non-colour of skin block, then block interior pixel point sets to 0;
7. judge whether that entire image has scanned? do not complete, then according to the sliding window moving step length of setting, moving window, then returns step 5, if image scanning completes, proceeds to step 8;
8. obtain final Face Detection result. Terminate.
Described step 1 comprises the following steps:
(1) (please refer to document " Jones; M.&Rehg; J.Statisticalcolormodelswithapplicationtoskindetection.I nternationalJournalofComputerVision; Springer; 2002,46,81-96 " from Compaq data set, download can be chained from author laboratory, belong to disclosed data set. ) in non-colour of skin sample in wherein, randomly draw the image block that is of a size of 10x10 and 20x20 size (the non-colour of skin number of blocks of the 10x10 of extraction be 384560 as non-colour of skin blocks of data collection; The non-colour of skin number of blocks of 20x20 extracted is 96140. );
(2) (please refer to document " Phung, S. from ECU data collection; Bouzerdoum, A.&Chai, D.SkinSegmentationUsingColorPixelClassification:Analysis andComparison.IEEETransactionsonPatternAnalysisandMachin eIntelligence, IEEEComputerSociety, 2005,27,148-154 "; download can be chained from author laboratory, belong to disclosed data set. ) in colour of skin sample in wherein, randomly draw the image block that is of a size of 10x10 and 20x20 size (the 10x10 colour of skin number of blocks of extraction be 188732 as colour of skin blocks of data collection; The 20x20 colour of skin number of blocks extracted is 47183. );
(3) (wherein the data centralization of 10x10 comprises colour of skin block 188732 and non-colour of skin block 384560 pieces the data set of 10x10 and 20x20 to be divided into 5 equal portions at random, equal timesharing, respectively colour of skin number of blocks is divided into 5 equal portions by sequence number ascending order, non-colour of skin number of blocks is also divided into 5 equal portions by sequence number ascending order, then using non-for first part of colour of skin block subset sums first part colour of skin block sub-combinations as the 1st equal portions, the similar combination of other equal portions), for training the cross validation of degree of depth conviction network.
Described step 2 comprises the following steps:
B1: (these are all the generic terms of this area, and in fact they are the name combinations of color space component, are used for representative color space title to select RGB, HSV, YCbCr and CIELab. Such as RGB tri-letters are red Red respectively, green Green and blue Blue. ) it is experiment color space, the extraordinary component that each channel components of often kind of color space is concentrated as feature, considers the CbCr color space after removing Y-component simultaneously, and experiment will detect the effect of depth model in above color space respectively.
B2: RGB+YCbCr is tested in color space combination respectively, RGB+HSV, RGB+CIELab, HSV+YCbCr, HSV+CIELab, YCbCr+CIELab, RGB+YCbCr+HSV, whole combination spaces in RGB+CIELab+HSV, YCbCr+CIELab+HSV and four kinds of spaces, test and combine effect spatially by detecting depth model respectively at these.
B3: the color space component extracting all pixels of moving window place image block, it is normalized to [0,1] scope, according to different colours space or combination, obtain set of eigenvectors and (first obtain all pixel coordinates in image block corresponding to window according to the window's position, then each pixel rgb value in region is obtained, difference according to the color space used, obtain the current vector value of each pixel and (specifically: if using rgb space, then directly obtain proper vector by connecting after R, G, B value normalization method of region interior pixel point; If using other color spaces, such as HSV, then the value being converted in HSV space by the rgb value of pixel, then obtains proper vector by connecting after H, S, V value normalization method; If the situation of color space combination, such as RGB+YCbCr, then six values of R, G, B, Y, Cb, the Cr after normalization method are connected to form proper vector).
Described step 3 comprises the following steps:
(1) determine that (this value is that experimentally result obtains to degree of depth conviction network packet, experiment display, if being set as 3 layers, then the accuracy rate of the model 20x10RGB model trained declines percentage point more than 5 containing 4 layers; If being set to 5 layers, the training time increases about 2��3 days under current computer hardware condition, but accuracy rate only promotes about 0.018 percentage point. So, balance efficiency and accuracy rate, adopt the numerical value of 4 layers. ), wherein hidden layer three layers, input layer one layer, by three limited Bohr hereby graceful machine form degree of depth conviction network;
(2) degree of depth skin color classifier increases by one layer of full interconnection network on degree of depth conviction network structure basis, for the output of classification results;
(3) sorter hidden layer structure (namely removing every layer of neuron number comprised outside input layer) is 100-100-100-2, and training minimum packets is 100, and iteration number of times is 200;
(4) on the data set of 10x10 and 20x20, corresponding degree of depth complexion model DSM is trained respectively10And DSM20, specific as follows:
Concrete execution step is as follows:
C4.1: obtain present image block feature x=(x0, x1..., xi);
C4.2: taking x=h (0) as degree of depth complexion model input layer, the h1 layer of training model;
C4.3: use h1 as the input data of the second layer;
C4.4: using h1 as learning sample, for training the h2 layer of RBM;
C4.5: iteration C4.3 and C4.4 step, instruct the training of all layers;
C4.6: using counterpropagation network adjusting and optimizing model parameter, wherein each parameter optimisation procedure is as follows:
C4.6.1: optimum color space and combination selection, adopt cross validation mode training and testing degree of depth complexion model respectively on the data set of 10x10 and 20x20, chooses optimal colors space according to accuracy rate, by mistake inspection rate etc., selects RGB color herein. Degree of depth skin color classifier performance under experiment test different colours spatial array simultaneously.
C4.6.2: degree of depth conviction network architecture parameters is optimized, i.e. hidden layer neuron quantity optimization. Setting neuronal quantity respectively is 100,200 and 500 experiments carrying out the data set of 20x20 under rgb space, and result illustrates that the increase of neuronal quantity can not promote classifier performance significantly. The present invention uses neuronal quantity to be 100.
C4.6.3: the determination of iteration number of times in optimization, setting iteration number of times respectively is 50,100 and 200, and under rgb space, the data set of 20x20 is tested, and result illustrates that increasing iteration number of times can make performance promote to some extent, but increases simultaneously and expend time in. The present invention adopts iteration number of times to be 200.
Described step 4 comprises the following steps:
(1) setting step-length step size respectively is 2 pixels, 5 pixels and 10 pixels;
(2) according to different step-length setting during mobile sliding window, moving window is moved with vertical direction in the horizontal direction, it is determined that current window position.
Specific as follows:
D1: setting moving window size is the tile size that degree of depth complexion model is corresponding;
D2: setting moving window moving step length is 2 pixels, 5 pixels and 10 pixels;
D3: obtain current image length to be detected and wide data;
D4: from the upper left corner, obtains current window scope, and namely when the coordinate row that moves ahead adds length of window: row+patch_size, when prostatitis, coordinate adds window width: col+patch_size;
D5: acquisition current image block content: patchi=img (row:row+patch_size-1, col:col+patch_size-1 :);
D6: horizontal direction and vertical direction are with step distance moving window, and repeating step D4 and D5, until row and col reaches image maximum length and width.
One embodiment of the present of invention is as follows:
Step S1, sets up colour of skin blocks of data collection and non-colour of skin blocks of data collection, and this data centralization stores the image block of 10x10 and 20x20 (pixel) size, is divided into colour of skin image block and non-colour of skin image block, sets up data set by following process:
A1: choose the non-colour of skin image pattern in a secondary Compaq database at random, randomly draw the image block being of a size of 10x10 (pixel) and 20x20 (pixel) size thereon to deposit as non-colour of skin block image pattern, set the image block that every width image chooses 10-20 at random.
A2: choose the mark image of the colour of skin in ECU data storehouse at random, in area of skin color, ((the mark image in ECU data storehouse remains area of skin color of human body part and other parts is erased to white background that (rgb value that white background and pixel are is all 255 in non-1 region, after normalization method, the value of the pixel in white background region is 1, and the value of the pixel of area of skin color is not 1). So, in image pixel value be 1 be non-skin pixel point, pixel value be not 1 be exactly skin pixel part. ) in randomly draw the image block being of a size of 10x10 and 20x20 size and separately deposit as skin patch image, set, according to actual area of skin color size, the colour of skin block sample number that every width image chooses.
A3: the data set of 10x10 and 20x20 is divided into 5 equal portions at random, for training the cross validation of degree of depth conviction network.
Step S2, according to color space and combination producing feature collection thereof, is undertaken by following process:
B1: select RGB, HSV, YCbCr and CIELab are experiment color space, the extraordinary component that each channel components of often kind of color space is concentrated as feature, considering the CbCr color space after removing Y-component, experiment will detect the effect of depth model in above color space respectively simultaneously.
B2: RGB+YCbCr is tested in color space combination respectively, RGB+HSV, RGB+CIELab, HSV+YCbCr, HSV+CIELab, YCbCr+CIELab, RGB+YCbCr+HSV, whole combination spaces in RGB+CIELab+HSV, YCbCr+CIELab+HSV and four kinds of spaces, test and combine effect spatially by detecting depth model respectively at these.
B3: the color space component extracting all pixels of moving window place image block, is normalized to [0,1] scope, according to different colours space or combination, obtains set of eigenvectors.
Step S3, arranges the structure of degree of depth conviction network, model structure and parameter, optimizes degree of depth complexion model parameter, trains and obtain optimal depth complexion model, and this step comprises following process:
C1: use limited Bohr hereby graceful machine (RBM) form degree of depth conviction network, wherein RBM comprises visible layer and hidden layer, and its joint probability distribution represents and isThe present invention sets and uses three layers of RBM to form degree of depth conviction network, and namely degree of depth conviction network packet is containing 4 layers, wherein hidden layer three layers, input layer one layer.
C2: build degree of depth skin color classifier, namely increases by one layer of full interconnection network, for the output of classification results on degree of depth conviction network structure basis. Wherein degree of depth conviction schematic network structure is as shown in Figure 2, and degree of depth complexion model schematic diagram is as shown in Figure 3.
C3: setting model parameter, wherein sorter hidden layer structure (namely removing every layer of neuron number comprised outside input layer) is 100-100-100-2, and training minimum packets is 100, and iteration number of times is 200;
C4: train degree of depth complexion model DSM on the data set of 10x10 and 20x20 respectively10And DSM20, concrete execution step is as follows:
C4.1: obtain present image block feature x=(x0, x1..., xi);
C4.2: taking x=h (0) as degree of depth complexion model input layer, the h1 layer of training model;
C4.3: use h1 as the input data of the second layer;
C4.4: using h1 as learning sample, for training the h2 layer of RBM;
C4.5: iteration C4.3 and C4.4 step, instruct the training of all layers;
C4.6: using counterpropagation network adjusting and optimizing model parameter, wherein each parameter optimisation procedure is as follows:
C4.6.1: optimum color space and combination selection, adopt cross validation mode training and testing degree of depth complexion model respectively on the data set of 10x10 and 20x20, chooses optimal colors space according to accuracy rate, by mistake inspection rate etc., selects RGB color herein. Degree of depth skin color classifier performance under experiment test different colours spatial array simultaneously.
C4.6.2: degree of depth conviction network architecture parameters is optimized, i.e. hidden layer neuron quantity optimization. Setting neuronal quantity respectively is 100,200 and 500 experiments carrying out the data set of 20x20 under rgb space, and result illustrates that the increase of neuronal quantity can not promote classifier performance significantly. The present invention uses neuronal quantity to be 100.
C4.6.3: the determination of iteration number of times in optimization, setting iteration number of times respectively is 50,100 and 200, and under rgb space, the data set of 20x20 is tested, and result illustrates that increasing iteration number of times can make performance promote to some extent, but increases simultaneously and expend time in. The present invention adopts iteration number of times to be 200.
Step S4, input test pattern picture, obtains image length, the master datas such as width.
Step S5, it may also be useful to moving window strategy, obtains tile location test pattern picture to be detected, and wherein moving window scanning image process is as follows:
D1: setting moving window size is the tile size that degree of depth complexion model is corresponding;
D2: setting moving window moving step length is 2 pixels, 5 pixels and 10 pixels;
D3: obtain current image length to be detected and wide data;
D4: from the upper left corner, obtains current window scope, and namely when the coordinate row that moves ahead adds length of window: row+patch_size, when prostatitis, coordinate adds window width: col+patch_size;
D5: acquisition current image block content: patchi=img (row:row+patch_size-1, col:col+patch_size-1 :);
D6: horizontal direction and vertical direction are with step distance moving window, and repeating step D4 and D5, until row and col reaches image maximum length and width.
Step S6, inputs the feature of moving window place image block the degree of depth conviction network trained, obtains the classification results of this image block, and this step comprises following process:
E1: if degree of depth skin color classifier Output rusults is colour of skin block, then block interior pixel point puts 1;
E2: otherwise be non-colour of skin block, pixel sets to 0;
Does step S7, judge that sliding window has scanned? namely judge whether following condition is set up:
Row+patch_size > height, col+patch_size > width
If condition is set up, then scan, submitted to and be false, go to step S5, obtained next video in window block and continue detection.
Step S8, completes the Face Detection of image based on moving window framework, obtains area of skin color of human body figure.
The present invention proposes a data set in units of image block, introduce the method for degree of depth study in Face Detection field first, namely use degree of depth conviction network to carry out the training and testing of skin color classifier; Based on moving window strategy, the sorter trained is used for the area of skin color segmentation of image; In the process of training degree of depth complexion model, consider have rated the classifier performance under four kinds of seperate color spaces and their different combined situation respectively; Have rated the parameter such as different neuronal quantities and iteration number of times to the impact of result simultaneously. Obtain the degree of depth complexion model of two kinds of yardsticks, based on the sliding window frame of different step-length, it is proposed to method can obtain Shandong rod, effectively area of skin color of human body segmentation result under complicated environment, be better than traditional method adding morphological operation by pixel classifications.
The present invention proposes a kind of taking image block as processing the skin color detection method based on degree of depth confidence network of unit. The method is taking image block as base conditioning unit, instead of taking discrete pixel as processing unit, just considers the space related information of pixel in detection, instead of re-uses morphological operation and carry out spatial manipulation after detection; Simultaneously, the inventive method proposes the degree of depth network structure for Face Detection first, based on the successful Application that degree of depth learning algorithm obtains at computer vision field, the degree of depth conviction network of training in units of image block, and adopt special moving window strategy and the model trained is used for area of skin color segmentation. Experimental result shows, the method that the present invention proposes can obtain area of skin color segmentation result more accurately under complex scene, it is better than the method merging morphological operation by pixel classifications commonly used, and by the performance reaching main stream approach in the performance of pixel classifications.
Technique scheme is one embodiment of the present invention, for those skilled in the art, on the basis that the present invention discloses application method and principle, it is easy to make various types of improvement or distortion, and it is not limited only to the method described by the above-mentioned embodiment of the present invention, therefore previously described mode is just preferred, and does not have restrictive meaning.
Claims (6)
1. the area of skin color of human body dividing method based on degree of depth conviction network, it is characterised in that: described method comprises:
(1). set up colour of skin blocks of data collection and non-colour of skin blocks of data collection;
(2). produce the degree of depth conviction network input feature vector collection in distinct colors space and combination thereof;
(3). the degree of depth conviction network of training different scale, and optimize and determine parameters;
(4). input test pattern picture, under different scale retrains, it may also be useful to moving window strategy obtains current image block scope to be detected;
(5). on the image block that current moving window is determined, extract color space component or its combination of needs, obtain input feature vector vector;
(6). input feature vector vector being inputted the degree of depth conviction network trained, obtains the classification results of image block, if classification results is colour of skin block, then block interior pixel point puts 1; If classification results is non-colour of skin block, then block interior pixel point sets to 0;
(7). judge whether entire image scanning completes, if not, then according to the sliding window moving step length moving window of setting, then return step (5), if it does, then proceed to step (8);
(8). obtain final Face Detection result, terminate;
Wherein, described step (2) comprises the following steps:
B1: select RGB, HSV, YCbCr and CIELab to be experiment color space, the extraordinary component that each channel components of often kind of color space is concentrated as feature, considers the CbCr color space after removing Y-component simultaneously;
B2: whole combination spaces that color space combines RGB+YCbCr, RGB+HSV, RGB+CIELab, HSV+YCbCr, HSV+CIELab, YCbCr+CIELab, RGB+YCbCr+HSV, RGB+CIELab+HSV and YCbCr+CIELab+HSV and four kinds of spaces are tested respectively, detection depth model combines effect spatially at these;
B3: the color space component extracting all pixels of moving window place image block, is normalized to [0,1] scope, according to different colours space or combination, obtains set of eigenvectors;
Described step (3) comprises the following steps:
(31) determine that degree of depth conviction network packet is containing 4 layers, wherein hidden layer three layers, input layer one layer, by three limited Bohr hereby graceful machine form degree of depth conviction network;
(32) degree of depth skin color classifier increases by one layer of full interconnection network on degree of depth conviction network structure basis, for the output of classification results;
(33) sorter hidden layer structure is 100-100-100-2, and training minimum packets is 100, and iteration number of times is 200;
(34) on the data set of 10x10 and 20x20, corresponding degree of depth complexion model DSM is trained respectively10And DSM20��
2. the area of skin color of human body dividing method based on degree of depth conviction network according to claim 1, it is characterised in that: described step (1) comprises the following steps:
(11) from the non-colour of skin sample of Compaq data centralization, the image block being of a size of 10x10 and 20x20 size is randomly drawed as non-colour of skin blocks of data collection;
(12) from the colour of skin sample that ECU data is concentrated, the image block being of a size of 10x10 and 20x20 size is randomly drawed as colour of skin blocks of data collection;
(13) data set of 10x10 and 20x20 is divided into 5 equal portions at random, for training the cross validation of degree of depth conviction network.
3. the area of skin color of human body dividing method based on degree of depth conviction network according to claim 1, it is characterised in that: described step (34) comprises the following steps:
C4.1: obtain present image block feature x=(x0, x1..., xi);
C4.2: taking x=h (0) as degree of depth complexion model input layer, the h1 layer of training model;
C4.3: use h1 as the input data of the second layer;
C4.4: using h1 as learning sample, for training the h2 layer of RBM;
C4.5: iteration C4.3 and C4.4 step, instruct the training of all layers;
C4.6: using counterpropagation network adjusting and optimizing model parameter, wherein each parameter optimisation procedure is as follows:
C4.6.1: optimum color space and combination selection: adopt cross validation mode training and testing degree of depth complexion model respectively on the data set of 10x10 and 20x20, optimal colors space is chosen, simultaneously the degree of depth skin color classifier performance under experiment test different colours spatial array according to accuracy rate, by mistake inspection rate.
C4.6.2: degree of depth conviction network architecture parameters is optimized, i.e. hidden layer neuron quantity optimization: setting neuronal quantity respectively is 100,200 and 500, carries out the experiment of the data set of 20x20 under optimal colors space, finds best neuronal quantity;
C4.6.3: the determination of iteration number of times in optimization: setting iteration number of times respectively is 50,100 and 200, and the data set of 20x20 is tested under optimal colors space, finds best iteration number of times.
4. the area of skin color of human body dividing method based on degree of depth conviction network according to claim 1, it is characterised in that: described step (4) comprises the following steps:
(41) setting step-length step size respectively is 2 pixels, 5 pixels and 10 pixels;
(42) according to different step-length setting during mobile sliding window, moving window is moved with vertical direction in the horizontal direction, it is determined that current window position.
5. the area of skin color of human body dividing method based on degree of depth conviction network according to claim 4, it is characterised in that: described step (42) comprises the following steps:
D1: setting moving window size is the tile size that degree of depth complexion model is corresponding;
D2: setting moving window moving step length is 2 pixels, 5 pixels and 10 pixels;
D3: obtain current image length to be detected and wide data;
D4: from the upper left corner, obtains current window scope, and namely when the coordinate row that moves ahead adds length of window: row+patch_size, when prostatitis, coordinate adds window width: col+patch_size;
D5: obtain current image block content:
Patchi=img (row:row+patch_size-1, col:col+patch_size-1 :);
D6: horizontal direction and vertical direction are with step distance moving window, and repeating step D4 and D5, until row and col reaches image maximum length and width.
6. the area of skin color of human body dividing method based on degree of depth conviction network according to claim 4, it is characterised in that: described step (5) is achieved in that
First determine moving window position, obtain the coordinate of the whole pixels in window area, then extract the channel value of each pixel in distinct colors space or spatial array, i.e. color space component values, morphogenesis characters component of connecting after normalization method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510497519.XA CN105138975B (en) | 2015-08-13 | 2015-08-13 | A kind of area of skin color of human body dividing method based on degree of depth conviction network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510497519.XA CN105138975B (en) | 2015-08-13 | 2015-08-13 | A kind of area of skin color of human body dividing method based on degree of depth conviction network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105138975A CN105138975A (en) | 2015-12-09 |
| CN105138975B true CN105138975B (en) | 2016-06-01 |
Family
ID=54724319
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510497519.XA Expired - Fee Related CN105138975B (en) | 2015-08-13 | 2015-08-13 | A kind of area of skin color of human body dividing method based on degree of depth conviction network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105138975B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105719303B (en) * | 2016-01-25 | 2018-04-17 | 杭州职业技术学院 | Magnetic resonance prostate 3D rendering dividing method based on more depth belief networks |
| CN106355151B (en) * | 2016-08-30 | 2019-10-01 | 电子科技大学 | A kind of three-dimensional S AR images steganalysis method based on depth confidence network |
| CN106960185B (en) * | 2017-03-10 | 2019-10-25 | 陕西师范大学 | The Pose-varied face recognition method of linear discriminant deepness belief network |
| CN108460362B (en) * | 2018-03-23 | 2021-11-30 | 成都品果科技有限公司 | System and method for detecting human body part |
| CN111681229B (en) * | 2020-06-10 | 2023-04-18 | 创新奇智(上海)科技有限公司 | Deep learning model training method, wearable clothes flaw identification method and wearable clothes flaw identification device |
| CN111815653B (en) * | 2020-07-08 | 2024-01-30 | 深圳市梦网视讯有限公司 | Method, system and equipment for segmenting human face and body skin color region |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103793718A (en) * | 2013-12-11 | 2014-05-14 | 台州学院 | Deep study-based facial expression recognition method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8055067B2 (en) * | 2007-01-18 | 2011-11-08 | DigitalOptics Corporation Europe Limited | Color segmentation |
| CN103400146B (en) * | 2013-07-19 | 2017-03-01 | 北京工业大学 | Chinese medicine complexion recognition method based on color modeling |
-
2015
- 2015-08-13 CN CN201510497519.XA patent/CN105138975B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103793718A (en) * | 2013-12-11 | 2014-05-14 | 台州学院 | Deep study-based facial expression recognition method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105138975A (en) | 2015-12-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105138975B (en) | A kind of area of skin color of human body dividing method based on degree of depth conviction network | |
| CN106203395B (en) | Face attribute recognition method based on multitask deep learning | |
| CN102054178B (en) | A kind of image of Chinese Painting recognition methods based on local semantic concept | |
| CN108280397B (en) | Human body image hair detection method based on deep convolutional neural network | |
| CN108537239B (en) | Method for detecting image saliency target | |
| CN108009509A (en) | Vehicle target detection method | |
| CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
| CN109344874A (en) | A kind of automatic chromosome analysis method and system based on deep learning | |
| CN103106265B (en) | Similar image sorting technique and system | |
| CN105893925A (en) | Human hand detection method based on complexion and device | |
| CN104408449B (en) | Intelligent mobile terminal scene literal processing method | |
| CN107977671A (en) | A kind of tongue picture sorting technique based on multitask convolutional neural networks | |
| CN107909015A (en) | Hyperspectral image classification method based on convolutional neural networks and empty spectrum information fusion | |
| CN107590489A (en) | Object detection method based on concatenated convolutional neutral net | |
| CN107403183A (en) | The intelligent scissor method that conformity goal is detected and image segmentation is integrated | |
| CN104573685B (en) | A kind of natural scene Method for text detection based on linear structure extraction | |
| CN108846446A (en) | The object detection method of full convolutional network is merged based on multipath dense feature | |
| CN104134071A (en) | Deformable part model object detection method based on color description | |
| CN108734719A (en) | Background automatic division method before a kind of lepidopterous insects image based on full convolutional neural networks | |
| CN111967313B (en) | Unmanned aerial vehicle image annotation method assisted by deep learning target detection algorithm | |
| CN108921120B (en) | Cigarette identification method suitable for wide retail scene | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN106203496B (en) | Hydrographic curve extracting method based on machine learning | |
| CN112906550B (en) | Static gesture recognition method based on watershed transformation | |
| CN105069816B (en) | A kind of method and system of inlet and outlet people flow rate statistical |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20160601 |