CN103793690A - Human body biotic living body detection method based on subcutaneous bloodstream detection and application - Google Patents

Abstract

The invention relates to a biotic identification living body detection method based on subcutaneous bloodstream detection. The colorful video, collected by common video collection equipment, of a skin area under normal light conditions serves as original signals, blood flowing signals contained in the original signals serve as a living body basis, and a deep learning neural network composed of RBM is used to carry out classification and identification on input signals. The method carries out living body detection on a sample based on the biotic characteristics of a living body, and is high in reliability. The method only uses the common video collection equipment which is necessary equipment for a biotic identification system, hardware equipment does not need to be additionally provided for the system, cost is low, and an algorithm is simple, easy to achieve and actual requirements on different occasions can be met.

Description

A kind of human-body biological biopsy method and application of surveying based on subcutaneous blood flow

Technical field

The present invention relates to technical field of biometric identification, relate to the method for the discrimination to human body live body in bio-identification, especially a kind of human-body biological biopsy method of surveying based on subcutaneous blood flow.

Background technology

Along with scientific and technological development and some technology ripe gradually in technical field of biometric identification in recent years, bio-identification (authentication) technology is widely used gradually.This technology refers to some biological characteristics based on human body, comprises that physiological characteristic and behavioural characteristic identify, differentiate a kind of technology of individual identity.This technology mainly relies on identification physiological characteristic in application aspect at present, generally includes recognition of face, refers to the identification of (palm) line, iris recognition etc.The uniqueness of biological characteristic just found by Chinese scholar as far back as the Tang Dynasty, and Modern West scholar also early has a large amount of research to this.And it is early stage really biological characteristic to be come across to the seventies in last century for the system of identification.In recent years, increasingly mature along with the decline of physical characteristics collecting equipment price and various recognition methodss, bio-identification (authentication) technology has incorporated in our daily life gradually, at numerous areas, as there is extensive application in the fields such as system login, customs's safety check, for example China has started citizen's fingerprint collecting has been stored in to the work in China second-generation identity card.

Bio-identification has the advantage of validity and convenience, but impersonation attack is a significant threat of biological recognition system always.For example, for face identification system, it is counterfeit that assailant can carry out identity with human face photo; And the rubber moulding that fingerprint, Palm Print Recognition System also may use silica gel or similar material the to make system of out-tricking.To this type of bogus attack, there are a lot of scholars to carry out research abroad, also delivered some documents and materials, prove to use these counterfeit means really can realize to a certain extent the attack to system.

Along with the application of biological identification technology is further extensive, security of system problem highlights day by day, more and more studied personnel's extensive concern.Wherein, a kind of thinking of head it off is, in obtaining human-body biological sample, will carry out live body detection to biological specimen simultaneously.Live body detection can guarantee obtaining liking biological living of biological information, thereby can resist the various biological specimens that copy, the such as attack to Verification System such as photo, plastic cement finger print.Become an important directions in field of biological recognition for the research of live body detection technique, in recent years, there are a large amount of research work and some important academic conferences both at home and abroad in this field, and have relevant work and paper publishing in the authoritative meeting of some association areas.Some conventional technology comprise physiological behavior (for example identifying action nictation of face), the environmental characteristic of live body and the physiological signal of live body (such as skin sweat, EEG level, thermal infrared characteristic etc.) based on live body at present.But these methods each defectiveness again, some need to be by the computing of large amount of complex, and some needs specific installation support, and also some method user experiences badly, is difficult to meet the application request of various complicated occasions, and ease for use and reliability aspect have much room for improvement.

By retrieval, not yet find the patent publication us relevant to patented claim of the present invention.

Summary of the invention

The object of the invention is to prevent bogus attack problem for solving in biological recognition system, a kind of method of human-body biological sample being carried out to live body detection has been proposed, utilize image processing techniques and degree of deep learning neural network to carry out live body detection to biological specimen, the method is with low cost, simple to operate, reliability is high, can meet the actual requirement of different occasions.

To achieve these goals, technical scheme of the present invention is as follows:

A kind of bio-identification biopsy method based on subcutaneous blood flow, take the color video of skin area under the normal illumination condition of ordinary video collecting device collection as original signal, take the blood flow signal that contains in original signal as live body foundation, and use the degree of deep learning neural network that formed by RBM in addition Classification and Identification of input signal.

And, the described bio-identification biopsy method based on subcutaneous blood flow, concrete steps are as follows:

(1) from colour-video signal, obtaining can category signal, and obtaining step is as follows successively:

A. use the color video collecting device of visible light frequency band to gather color video or the coherent image picture of skin area, sample frequency is higher than more than 8fps;

B. from video or coherent image, intercept out the sequence of certain time length;

C. using the region that comprises skin in image as ROI, with ROI region, extract the red, green, blue triple channel pixel grey scale mean value in ROI region, the vector of composition 3 × N is as raw data; Wherein, the frame number that N is video, is greater than 50;

D. raw data is carried out to elimination trend fluctuation and process, obtaining can category signal;

(2) can use degree of deep learning neural network carry out discriminant classification by category signal to (1) described in step, classification step be as follows successively:

A. the training of degree of deep learning neural network:

1. use in step (1) described acquisition methods that can category signal, from human body live body, collection signal is as the positive sample of network training;

2. use in step (1) described acquisition methods that can category signal, from non-human live body or prosthese, collection signal is as the negative sample of network training;

3. every one deck RBM of network is carried out to unsupervised pre-training;

4. use gradient descent algorithm, the tagged positive sample of network input, negative sample are trained; ;

B. use degree of deep learning neural network to carry out live body detection:

1. use in step (1) described acquisition methods that can category signal, on sample, gather can category signal detecting;

2. in using in step (2), the training of a degree of deep learning neural network is to carrying out identification and classification by category signal, whether thereby realizing is the differentiation of live body to sample, if degree of deep learning neural network is finally output as positive sample, be finally output as one and be greater than 0 positive number, think and can gather from human body biopsy sample by category signal; If degree of deep learning neural network is finally output as negative sample, be finally output as one and be less than 0 negative, think category signal sample to be non-human live body.

And, described step (1) in a colored video capture device be common camera, camera or industrial camera.

And, described step (1) in b the duration of sequence be 3-5 second.

And (1) described step is carried out the fluctuation of elimination trend to raw data in d and processed and use SPA method to go trend fluctuation to process, step is as follows:

If raw data is X, going data after trending is Y, uses formula:

$Y=(I-{(I+{\mathrm{\λ}}^2{D}_2^T{D}_2)}^{-1})X$

Wherein, I is N rank unit matrixs, and D2 is N-2 × N rank matrixes, and concrete form is as follows:

N is signal length, and its span is for being greater than 50; λ can control the cut frequency of trending, and the span of λ is 10～20;

And, described step (2) a every one deck RBM of network to be carried out to the concrete steps of training of unsupervised pre-training in 3. as follows:

RBM establishes two layers altogether, and one is input layer, and another is output layer, and this layer can be regarded as the feature extraction layer to input, the cascade form that each RBM is input-output;

Before training, first the training set random alignment again that 500 samples form will at least respectively be selected to be no less than by positive sample, negative sample, then by all sample mean groupings;

In training, input mono-group of data of RBM rather than data at every turn;

If the input layer of RBM network has i node, output layer has j node, and input layer is biased to a={a _i, output layer is biased to b={b _i, interlayer node weights are w={w _ij; The v that is input as that makes RBM, is output as h, and excitation function is sigmoid type;

The pre-training process of single RBM is as follows:

1. use random number initializes weights, to a, b, w random number initialization;

2. under current network parameter, calculate the output of one group of input data v, making it is h (0);

3. backwards calculation RBM, uses h (0) reconstruct input, and making it is v (1);

4. take v (1) as input, calculate output under current network parameter, making it is h (1);

5. calculate output expectation and reconstruct output expectation.Output expects to be v'*h (0), and reconstruct output is expected for v (1) ' * h (1);

6. upgrade network parameter a by following formula, b, w:

$\mathrm{\Δa}=m*\mathrm{\Δa}+\left(\frac{\∈}{100}\right)*(\underset{100}{\mathrm{\Σ}}v-\underset{100}{\mathrm{\Σ}}v\left(1\right))$

$\mathrm{\Δb}=m*\mathrm{\Δb}+\left(\frac{\∈}{100}\right)*(\underset{100}{\mathrm{\Σ}}h\left(0\right)-\underset{100}{\mathrm{\Σ}}h\left(1\right))$

$\mathrm{\Δw}=m*\mathrm{\Δw}+\∈*\left(\frac{v^{\′}*h\left(0\right)-{v\left(1\right)}^{\′}*h\left(1\right)}{100}\right)$

Wherein, m and ∈, for training coefficient, generally get m=0.8～0.9, ∈=0.00005～0.00015;

7. to 200 groups of all Data duplications 2.-6.;

8. 2.-7. 50 iteration is carried out time, and RBM trains end in advance;

Training after ground floor RBM according to step above, by all training data input ground floor RBM, exported the input as second layer RBM, the lower one deck RBM of training, by that analogy, the training successively of each layer, until last one deck RBM.

And the number of plies of described RBM is 1～8 layer.

And, the described step concrete steps that (2) a carries out unsupervised pre-training to every one deck RBM of network in are 3. as follows: described degree of deep learning neural network is made up of 4 RBM, these 4 RBM called after RBM_in respectively, RBM_h1, RBM_h2, RBM_out, 4 cascade forms that RBM is input-output;

RBM_in is the input layer of network, and what it was input as 600 dimensions can category signal, and output node is 5000; RBM_h1 is first hidden layer of network, and it is input as the output of RBM_in, and totally 5000 nodes, are output as 1000 nodes; RBM_h2 is second output layer of network, and it is input as the output of RBM_h1, and totally 1000 nodes, are output as 100 nodes; RBM_out is the output layer of network, and it is input as the output of RBM_h2, and totally 100 nodes, are output as 1 node;

The excitation function of all nodes is all selected sigmoid type; In training process, first use not tagged training set to train in advance one by one 4 RBM, the object of pre-training is to allow RBM matching reconstruct data better;

First RBM_in is trained in advance, reach after training objective at it, export as training sample using it, training RBM_h1, trains 4 all RBM by that analogy.

And described step (2) a is used the tagged positive negative sample concrete steps of training as follows to network in 4.:

1. the quantity of described RBM is 4, uses the RBM network consisting training, and the degree of deep learning neural network of composition is of five storeys altogether, and every node layer is respectively 200,5000,1000,100,1, and the weights of each layer and biasing are used the training result of pre-training stage;

2. training sample tagged and again divide into groups, every 1000 data being divided into one group, totally 20 groups of data;

3. use training sample to train above-mentioned 5 layer networks, training method is used basic gradient descent algorithm.

The application of bio-identification biopsy method based on subcutaneous blood flow as above aspect technical field of biometric identification.

Advantage of the present invention and good effect are:

The inventive method is passed through (interested region, ROI region in color video/image sequence, RegionofInterest) the equal value sequence of RGB passage pixel goes trend fluctuation to process, therefrom extracting can category signal, and this signal is due to the pulse information that has comprised live body; Utilize by multiple RBM(RestrictedBoltzmannMachines, limited Boltzmann machine) form network signal is classified, thereby whether judgment signal picks up from live body.The Biological characteristics of the method based on live body carries out live body detection to sample, and reliability is high; Method is only used common video capture device, and these equipment are the essential equipment in biological recognition system, need not additionally add hardware device to system, with low cost, and algorithm is simply easy to realize, and can meet the actual requirement of different occasions.

Accompanying drawing explanation

Fig. 1 is structure principle chart of the present invention;

Fig. 2 be in the present invention ROI region choose sample graph; Wherein, a is biopsy sample figure, and b is prosthese (picture) sample graph;

Fig. 3 is degree of deep learning neural network structural drawing of the present invention;

Fig. 4 is original ROI of the present invention region RGB passage average figure; Wherein, top figure be from live body (Fig. 2 a) extract original signal, below figure be from prosthese (Fig. 2 b) extraction original signal;

Fig. 5 signal graph going after trending of the present invention; Wherein, top curve is for from live body, (signal that Fig. 2 a) obtains, lower curve is from the prosthese (signal that Fig. 2 b) obtains.

Embodiment

Below in conjunction with embodiment, the present invention is further described; Following embodiment is illustrative, is not determinate, can not limit protection scope of the present invention with following embodiment.

The bio-identification biopsy method that the present invention is based on blood flow, its basic ideas are: the pulse information that in colour-video signal, the gray-scale value of the R in skin of living body region, G, tri-passages of B has comprised live body, and prosthese is not contain this information.Along with beat (heartbeat) of people's pulse, the RGB passage gray scale of the colour of skin of human body skin can produce slight variation along with blood flow, although changing human eye, this is difficult to catch, but can collect the signal of this variation by common color video capture device, the red, green, blue passage of original signal be done respectively to the fluctuation of elimination trend and process; In data after treatment, comprise the mobile signal (heartbeat) of blood of human body, use one by 4 grades of RBM(RestrictedBoltzmannMachines, limited Boltzmann machine) composition degree of deep learning neural network these data are trained, obtain the network structure that can distinguish live body and non-living body, utilize this network structure can judge whether input data are biological living.Pick up from respectively the vision signal of live body and non-living body by contrast, can distinguish live body and non-living body.The video that contains pulse information is considered to pick up from live body, that is to say and can judge that current detection sample is live body.

The present invention is based on the human-body biological biopsy method that subcutaneous blood flow is surveyed, its concrete thought is as follows:

As shown in Figure 1, whether, by obtaining the color video frequency image of skin area in sample to be tested, extracting effectively can category signal, use the degree of deep learning neural network being made up of multilayer RBM to classify to sample, thereby be that live body judges to it.

The described color video frequency image that obtains skin area in sample to be tested, extract effectively can category signal concrete steps as follows:

A. from be not less than the color video of 8fps or image sequence, intercept the signal of certain time length, generally get 5-8 second.Because the scope of live body human heart rate is roughly 0.7～4Hz(per minute 40 times～240 times) between, according to nyquist sampling theorem, find that the Least sampling rate of video or image sequence is 8fps.In fact, most of video capture devices, as first-class in common shooting, sampling rate generally will be higher than 15fps.In description below, one to show color video 200 two field pictures that intercept 30fps be example.

B. from video or image sequence, selecting the region that comprises skin is the interested region of ROI(, RegionofInterest).For example, in recognition of face, delimiting the human face region being detected is ROI, in palmmprint, fingerprint recognition, take the skin area of corresponding palm, finger part as ROI.

C. get the three-channel gray average of red, green, blue in each two field picture of ROI region, form a vectorial X, that is:

X=[X ₁，X _gX _b]???(1）

Wherein, x _r, x _g, x _bbe respectively the sequence of 200 frame ROI region RGB triple channel gray averages, the one-dimensional vector that namely length is 200.

D. respectively to x _r, x _g, x _buse SPA method (SmoothnessPriorsApproach) to go trend fluctuation to process, the trending removing in signal disturbs, and obtains:

Y=[y ₁，y _gy _b]???(2）

In (2), can observe, the signal frequency of green, blue two passages and the degree of correlation of pulse frequency are very large.

The vectorial Y finally obtaining is effectively can category signal.

Described use is by multilayer RBM(RestrictedBoltzmannMachines, limited Boltzmann machine) whether the degree of deep learning neural network of composition classify to sample, thereby be that live body judges that concrete steps are as follows to it:

A. described network forms by 4 layers, and every layer is made up of a RBM, called after RBM_in respectively, RBM_h1, RBM_h2, RBM_out.Wherein, what being input as of RBM_in obtained above can category signal, and it is input as three-channel original 200 frame data of RGB, totally 600 inputs, and the output node of RBM_out is 1.All the other inputs of each layer are the output on upper strata.The input-output nodes of each RBM is respectively: 600-5000(RBM_in), 5000-1000(RBM_h1), 1000-100(RBM_h2), 100-1(RBM_out).

B. using the method introduced to go up extraction at real live body and prosthese (photo, plastic mold etc.) respectively above can category signal, successively 4 RBM is trained in advance as positive negative sample.So-called pre-training is a kind of unsupervised learning, its objective is and allows each RBM have better matching and re-configurability to its input signal, to solve the local extremum problem existing in neural metwork training.In the pre-training stage, positive negative sample mixes formation training sample, and without this label of application of sample.Here, positive negative sample is respectively selected 10000.

C. on the basis of pre-training, use tagged positive negative sample to carry out the fine setting under supervised learning condition to entire depth network.After fine setting, the output node of network can carry out to category signal the discriminant classification of live body/prosthese.

By reference to the accompanying drawings the present invention is described in detail now, accompanying drawing illustrates basic structure of the present invention in a schematic way, therefore only shows the formation relevant with the present invention.In case below, detect as example take live body in face identification system, frame frequency is 30fps, selects continuous 200 two field picture composition original sample videos.

Embodiment 1

A bio-identification biopsy method based on blood flow, step is as follows:

(1) extraction that can category signal

First by using correlation method to determine the position of sample in video in original video, the technology such as the such as face detection in complex background.From having located the video of sample position, choose the skin area corresponding with sample as ROI, as shown in Figure 2.In each two field picture, get the gray-scale value of image RGB passage, video is converted to corresponding floating type Serial No..Extract regular length Serial No. (200 frame) form original signal, as formula 1. as described in.Go trend fluctuation to process original signal, obtaining can category signal.Final can category signal totally 600 dimensions.

Said extracted can category signal concrete steps as follows:

1, in video, identifying object is just located, and detects as first carried out face in complex background in recognition of face.

2, in the identifying object of location, select ROI region.The principle of selecting is to allow ROI region only comprise skin as far as possible, or maximizes the ratio of skin area in ROI as far as possible.For example, in recognition of face, the result that face detects has often comprised hair, and the background of face both sides, so should suitably inwardly shrink in delimiting ROI on human face region basis.In Fig. 2, in live body a and prosthese b sample, the green frame in outside is the human face region of identification, and the green frame in inner side is the ROI region after shrinking in proportion.

3, in each two field picture, the RGB passage average gray in ROI region is calculated, formation sequence in 200 continuous frames, original signal is as shown in Figure 4.In Fig. 4, upper figure be from live body (Fig. 2 a) extract original signal, figure below be from prosthese (Fig. 2 b) extraction original signal.

4, original signal is gone trend fluctuation process, the internal and external factor that removes various complexity affects the low-frequency fluctuation bringing.That here use is SPA(SmoothnessPriorsApproach, priori smoothing method) method, respectively the gray average sequence of three passages of RGB is processed.

If original signal is X, removing signal after trending is Y, uses formula:

$Y=(I-{(I+{\mathrm{\λ}}^2{D}_2^T{D}_2)}^{-1})X---\left(3\right)$

Wherein, I is N rank unit matrixs, and D2 is N-2 × N rank matrixes, and concrete form is as follows:

N is signal length, is here N=200.λ can control the cut frequency of trending, selects λ=20 here, and corresponding cut frequency is 0.6Hz.Remove trending signal Y after treatment as shown in Figure 5.In Fig. 5, top curve is for from live body, (signal that Fig. 2 a) obtains, lower curve is for from prosthese, (signal that Fig. 2 b) obtains, can find out live body signal, and especially turquoise two passages of signal have the feature with pulse wave clearly.

(2) can use degree of deep learning neural network to carry out discriminant classification by category signal:

The training of degree of deep learning neural network and use degree of deep learning neural network carry out live body detection:

Gather respectively each 10000 of positive negative sample, composition training set, with this training set training degree of deep learning neural network.

The selected network structure of the present invention is: the network being formed by 4 RBM, and called after RBM_in respectively, RBM_h1, RBM_h2, RBM_out, as shown in Figure 3.RBM_in is the input layer of network, and what it was input as 600 dimensions can category signal, and output node is 5000; RBM_h1 is first hidden layer of network, and it is input as the output of RBM_in, and totally 5000 nodes, are output as 1000 nodes; RBM_h2 is second hidden layer of network, and it is input as the output of RBM_h1, and totally 1000 nodes, are output as 100 nodes; RBM_out is the output layer of network, and it is input as the output of RBM_h2, and totally 100 nodes, are output as 1 node.The excitation function of all nodes is all selected sigmoid type.In training process, first use not tagged training set to train in advance one by one 4 RBM, the object of pre-training is to allow better matching reconstruct data of RBM.Due to 4 RBM cascade form that is input-output, therefore first RBM_in is trained in advance, reach after training objective at it, export as training sample using it, training RBM_h1, train by that analogy 4 all RBM, after pre-training finishes, the degree of deep learning neural network training should reflect the useful feature in original signal well.On this basis, training data is added to the label of positive negative sample, continue network to finely tune, the network after fine setting can well be to classifying by category signal, thus identification live body.

Above-mentioned training degree of deep learning neural network, and with this Network Recognition can category signal so that the degree of depth network structure of carrying out live body detection as shown in Figure 3, training and identification concrete steps are as follows:

1, the pre-training of each layer of RBM:

The present invention uses 4 RBM composition degree of depth networks.RBM(RestrictedBoltzmannMachines) be a kind of non-directed graph model, have two layers, one is input layer (or claiming visual layers), and another is hidden layer (output layer of RBM), and this layer can be regarded as the feature extraction layer to input.Before training, first will, by positive sample (live body), each 10000 training sets that form of negative sample (prosthese) random alignment again, then be divided into one group totally 200 groups, 100 samples.In training, input mono-group of data of RBM rather than data at every turn, for the situation of large data sample, grouping can obtain better training effect like this.If the input layer of RBM network has i node, output layer has j node, and input layer is biased to a={a _i, output layer is biased to b={b _i, interlayer node weights are w={w _ij.The v that is input as that makes RBM, is output as h, and excitation function is sigmoid type (S type).The pre-training process of single RBM is as follows:

1. use random number initializes weights, to a, b, w random number initialization.

2. under current network parameter, calculate the output of one group of input data v, making it is h (0).

3. backwards calculation RBM, uses h (0) reconstruct input, and making it is v (1).

4. take v (1) as input, calculate output under current network parameter, making it is h (1).

5. calculate output expectation and reconstruct output expectation.Output expects that, for v'*h (0), reconstruct output is expected for v (1) ' * h (1).

6. upgrade network parameter a by following formula, b, w:

$\mathrm{\Δa}=m*\mathrm{\Δa}+\left(\frac{\∈}{100}\right)*(\underset{100}{\mathrm{\Σ}}v-\underset{100}{\mathrm{\Σ}}v\left(1\right))---\left(4\right)$

$\mathrm{\Δb}=m*\mathrm{\Δb}+\left(\frac{\∈}{100}\right)*(\underset{100}{\mathrm{\Σ}}h\left(0\right)-\underset{100}{\mathrm{\Σ}}h\left(1\right))---\left(5\right)$

$\mathrm{\Δw}=m*\mathrm{\Δw}+\∈*\left(\frac{v^{\′}*h\left(0\right)-{v\left(1\right)}^{\′}*h\left(1\right)}{100}\right)---\left(6\right)$

Wherein, m and ∈, for training coefficient, generally get m=0.8, ∈=0.0001.

7. to 200 groups of all Data duplications 2.-6..

8. 2.-7. 50 iteration is carried out time, and RBM trains end in advance.

Training after RBM_in according to step above, by all training data input RBM_in, exported the input as RBM_h1, training RBM_h1, uses similar method training RBM_h2, RBM_out.

3, use tagged training sample, neural network is finely tuned to training, concrete steps are as follows:

1. use 4 RBM network consistings that train above, the degree of depth network of composition is of five storeys altogether, and every node layer is respectively 200,5000,1000,100,1, and the weights of each layer and biasing are used the training result of pre-training stage.

2. training sample added positive and negative samples label (positive sample label is 1, and negative sample label is-1) and again divided into groups, every 1000 data being divided into one group, totally 20 groups of data.

3. use training sample to train above-mentioned 5 layer networks, training method is used basic gradient descent algorithm.

4, use degree of depth network to carry out live body detection to recognition sample, concrete steps are as follows:

1. in video, recognition sample is carried out to just location.

2. according to the method for introducing, in selective recognition sample, comprise the ROI region of skin above.

3. extract the RGB passage gray average in ROI region in continuous 200 frame pictures, form original signal.

4. use formula (3), from original signal, extracting can category signal.

5. the degree of deep learning neural network can category signal input training, the output of degree of deep learning neural network is live body testing result, as the output of degree of deep learning neural network is judged to be positive sample (be finally output as and be greater than 0 positive number), think that this sample is human body biopsy sample, otherwise (be finally output as one and be less than 0 negative) is prosthese sample.

In the present embodiment, using the sample in Fig. 2 a as input, the final output valve of network is 0.982, and using the sample in Fig. 2 b as input, the final output valve of network is-1.011, and therefore whether be biological living in the present invention if can accurately differentiate sample.

Embodiment 2

A human-body biological biopsy method of surveying based on subcutaneous blood flow, step is as follows:

Change embodiment 1 the inside video sampling frequency is 15fps(15 frame/second), change sample frame number into 100, the input node of ground floor RBM in embodiment 1 is changed to 100, and other use network structure and the method identical with embodiment 1, can realize same recognition effect.

Embodiment 3

A human-body biological biopsy method of surveying based on subcutaneous blood flow, step is as follows:

Use video sampling frequency and the detection method identical with embodiment, the network structure of change embodiment 1 the inside is 6 layers, and each layer of Inport And Outport Node is respectively: ground floor RBM, 200 input nodes, 5000 output nodes; Second layer RBM, 5000 input nodes, 2000 output nodes; The 3rd layer of RBM, 2000 input nodes, 200 output nodes; The 4th layer of RBM, 200 input nodes, 50 output nodes; Layer 5 RBM, 50 input nodes, 10 output nodes; Layer 6 RBM, 10 input nodes, 1 output node.Can realize equally the recognition effect identical with embodiment 1.

Take above-mentioned foundation desirable case study on implementation of the present invention as enlightenment, by above-mentioned description, relevant staff can, not departing from the scope of this invention technological thought, carry out various change and modification completely.The technical scope of expecting to invent is not limited to the content on instructions, must determine its technical scope according to claim scope.

Claims (10)

1. the bio-identification biopsy method based on subcutaneous blood flow, it is characterized in that: take the color video of skin area under the normal illumination condition of ordinary video collecting device collection as original signal, take the blood flow signal that contains in original signal as live body foundation, and use the degree of deep learning neural network that formed by RBM in addition Classification and Identification of input signal.

2. the bio-identification biopsy method based on subcutaneous blood flow according to claim 1, is characterized in that: concrete steps are as follows:

(1) from colour-video signal, obtaining can category signal, and obtaining step is as follows successively:

A. use the color video collecting device of visible light frequency band to gather color video or the coherent image picture of skin area, sample frequency is higher than more than 8fps;

B. from video or coherent image, intercept out the sequence of certain time length;

C. using the region that comprises skin in image as ROI, with ROI region, extract the red, green, blue triple channel pixel grey scale mean value in ROI region, the vector of composition 3 × N is as raw data; Wherein, the frame number that N is video, is greater than 50;

D. raw data is carried out to elimination trend fluctuation and process, obtaining can category signal;

(2) can use degree of deep learning neural network carry out discriminant classification by category signal to (1) described in step, classification step be as follows successively:

A. the training of degree of deep learning neural network:

1. use in step (1) described acquisition methods that can category signal, from human body live body, collection signal is as the positive sample of network training;

2. use in step (1) described acquisition methods that can category signal, from non-human live body or prosthese, collection signal is as the negative sample of network training;

3. every one deck RBM of network is carried out to unsupervised pre-training;

4. use gradient descent algorithm, the tagged positive sample of network input, negative sample are trained; ;

B. use degree of deep learning neural network to carry out live body detection:

1. use in step (1) described acquisition methods that can category signal, on sample, gather can category signal detecting;

2. in using in step (2), the training of a degree of deep learning neural network is to carrying out identification and classification by category signal, whether thereby realizing is the differentiation of live body to sample, if degree of deep learning neural network is finally output as positive sample, be finally output as one and be greater than 0 positive number, think and can gather from human body biopsy sample by category signal; If degree of deep learning neural network is finally output as negative sample, be finally output as one and be less than 0 negative, think category signal sample to be non-human live body.

3. the bio-identification biopsy method based on subcutaneous blood flow according to claim 2, is characterized in that: described step (1) in a colored video capture device be common camera, camera or industrial camera.

4. the bio-identification biopsy method based on subcutaneous blood flow according to claim 2, is characterized in that: described step (1) in b the duration of sequence be 3-5 second.

5. the bio-identification biopsy method based on subcutaneous blood flow according to claim 2, is characterized in that: (1) described step is carried out the fluctuation of elimination trend to raw data in d and processed and use SPA method to go trend fluctuation to process, and step is as follows:

If raw data is X, going data after trending is Y, uses formula:

Wherein, I is N rank unit matrixs, and D2 is N-2 × N rank matrixes, and concrete form is as follows:

N is signal length, and its span is for being greater than 50; λ can control the cut frequency of trending, and the span of λ is 10～20.

6. the bio-identification biopsy method based on subcutaneous blood flow according to claim 2, is characterized in that: described step (2) a every one deck RBM of network to be carried out to the concrete steps of training of unsupervised pre-training in 3. as follows:

RBM establishes two layers altogether, and one is input layer, and another is output layer, and this layer can be regarded as the feature extraction layer to input, the cascade form that each RBM is input-output;

Before training, first the training set random alignment again that 500 samples form will at least respectively be selected to be no less than by positive sample, negative sample, then by all sample mean groupings;

In training, input mono-group of data of RBM rather than data at every turn;

If the input layer of RBM network has i node, output layer has j node, and input layer is biased to a={a _i, output layer is biased to b={b _i, interlayer node weights are w={w _ij; The v that is input as that makes RBM, is output as h, and excitation function is sigmoid type;

The pre-training process of single RBM is as follows:

1. use random number initializes weights, to a, b, w random number initialization;

2. under current network parameter, calculate the output of one group of input data v, making it is h (0);

3. backwards calculation RBM, uses h (0) reconstruct input, and making it is v (1);

4. take v (1) as input, calculate output under current network parameter, making it is h (1);

5. calculate output expectation and reconstruct output expectation.Output expects to be v'*h (0), and reconstruct output is expected for v (1) ' * h (1);

6. upgrade network parameter a by following formula, b, w:

Wherein, m and ∈, for training coefficient, generally get m=0.8～0.9, ∈=0.00005～0.00015;

7. to 200 groups of all Data duplications 2.-6.;

8. 2.-7. 50 iteration is carried out time, and RBM trains end in advance;

Training after ground floor RBM according to step above, by all training data input ground floor RBM, exported the input as second layer RBM, the lower one deck RBM of training, by that analogy, the training successively of each layer, until last one deck RBM.

7. the bio-identification biopsy method based on subcutaneous blood flow according to claim 6, is characterized in that: the number of plies of described RBM is 1～8 layer.

8. the bio-identification biopsy method based on subcutaneous blood flow according to claim 6, it is characterized in that: the described step concrete steps that (2) a carries out unsupervised pre-training to every one deck RBM of network in are 3. as follows: described degree of deep learning neural network is made up of 4 RBM, these 4 RBM called after RBM_in respectively, RBM_h1, RBM_h2, RBM_out, 4 cascade forms that RBM is input-output;

RBM_in is the input layer of network, and what it was input as 600 dimensions can category signal, and output node is 5000; RBM_h1 is first hidden layer of network, and it is input as the output of RBM_in, and totally 5000 nodes, are output as 1000 nodes; RBM_h2 is second output layer of network, and it is input as the output of RBM_h1, and totally 1000 nodes, are output as 100 nodes; RBM_out is the output layer of network, and it is input as the output of RBM_h2, and totally 100 nodes, are output as 1 node;

The excitation function of all nodes is all selected sigmoid type; In training process, first use not tagged training set to train in advance one by one 4 RBM, the object of pre-training is to allow RBM matching reconstruct data better;

First RBM_in is trained in advance, reach after training objective at it, export as training sample using it, training RBM_h1, trains 4 all RBM by that analogy.

9. according to the bio-identification biopsy method based on subcutaneous blood flow described in claim 2 or 8, it is characterized in that: described step (2) a is used the tagged positive negative sample concrete steps of training as follows to network in 4.:

1. the quantity of described RBM is 4, uses the RBM network consisting training, and the degree of deep learning neural network of composition is of five storeys altogether, and every node layer is respectively 200,5000,1000,100,1, and the weights of each layer and biasing are used the training result of pre-training stage;

2. training sample tagged and again divide into groups, every 1000 data being divided into one group, totally 20 groups of data;

3. use training sample to train above-mentioned 5 layer networks, training method is used basic gradient descent algorithm.

10. the bio-identification biopsy method based on subcutaneous blood flow as described in claim 1 to 9 any one is in the application aspect technical field of biometric identification.

Images