CN107832684A - A kind of intelligent vein authentication method and system with independent learning ability - Google Patents

A kind of intelligent vein authentication method and system with independent learning ability Download PDF

Info

Publication number
CN107832684A
CN107832684A CN201711013178.XA CN201711013178A CN107832684A CN 107832684 A CN107832684 A CN 107832684A CN 201711013178 A CN201711013178 A CN 201711013178A CN 107832684 A CN107832684 A CN 107832684A
Authority
CN
China
Prior art keywords
mrow
mtd
training
image
msub
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.)
Granted
Application number
CN201711013178.XA
Other languages
Chinese (zh)
Other versions
CN107832684B (en
Inventor
许炎
李稚春
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tonghua Technology Dalian Co ltd
Original Assignee
Tonghua Technology Dalian Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tonghua Technology Dalian Co ltd filed Critical Tonghua Technology Dalian Co ltd
Priority to CN201711013178.XA priority Critical patent/CN107832684B/en
Publication of CN107832684A publication Critical patent/CN107832684A/en
Application granted granted Critical
Publication of CN107832684B publication Critical patent/CN107832684B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/10—Image acquisition
    • G06V10/12—Details of acquisition arrangements; Constructional details thereof
    • G06V10/14—Optical characteristics of the device performing the acquisition or on the illumination arrangements
    • G06V10/147—Details of sensors, e.g. sensor lenses
    • 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
    • G06V40/14—Vascular patterns

Abstract

The invention discloses a kind of intelligent vein authentication method and system with independent learning ability, in method：Full convolutional neural networks deep learning model is carried out repeatedly, regularly global training and irregular local training are combined mode and carry out model training, before full convolutional neural networks deep learning model on-line running, global training is carried out once using the sample accumulated, and preserves feature extraction, feature recognition and the tagsort parameter of optimization；After model on-line running, the irregular local optimization trained, complete tagsort parameter is carried out to the model using the sample newly registered；Under given conditions, regularly global train is carried out to model using the sample newly accumulated, feature extraction, feature recognition and tagsort parameter are optimized again, ensure that the model turns into a kind of intelligent authentication system with independent learning ability while enhances the precision of vena identification.

Description

A kind of intelligent vein authentication method and system with independent learning ability
Technical field
The present invention relates to image recognition and processing technology field, more particularly to a kind of intelligence with independent learning ability are quiet Arteries and veins authentication method and system
Background technology
In recent years, as the fast development of biological identification technology, finger print identifying mode have been widely used.Simultaneously Iris, sound, face, vena metacarpea with higher confidential nature, refer to the authentication modes such as vein and be also more and more used.Its Middle fin- ger vein authentication mode is obtained due to the advantages that its confidentiality is high, not reproducible, certification is convenient, authenticating device small volume More it is widely applied.Finger vena identification generally uses finger figure of the various image segmentation algorithms from collection in the prior art Vein image is separated as in, forms the bianry image of vein train of thought, feature extraction algorithm is recycled, is carried from train of thought binary map Individual vein pattern is taken, finally certification sample is compared using feature recognition algorithms and registers the feature of sample, obtain authentication result. The authentication method of this algorithm of solidification in advance is primarily present that certification accuracy beyond algorithm is relatively low, complicated calculation in actual applications The problems such as method causes system cost higher hardware requirement height.Deep learning algorithm is also present in finger vena identification application at present Following problem, (1) reply dynamic cataloging method deficiency.Deep learning generally passes through supervised learning method using training data Training pattern.But the sample of classification to be identified is that dynamic is increased in practical application, supervised learning mistake before online implementing is only used Caused parameter in journey, it is impossible to identify all identification accuracys for increasing sample to be identified newly, influenceing new samples exactly.(2) it is right Learning data quantitative requirement is high, learning time length.Deep learning model is more complicated, and parameter optimization needs a large amount of training datas to make Support, while also need to successive ignition to complete to train.But finger vena sample data had not both had standard exercise number at present According to collection, also without sufficient amount of training data.(3) certification accuracy does not reach precision level also.Deep learning is to be based on mould Anthropomorphic class neuron system, the method for the purpose of substituting and be accomplished manually study and work, its recognition accuracy can exceed people Class reaches level of percent, but to reach the level of this precision permillage of authenticating device, particularly field of identity authentication pair False acceptance rate FAR (False Acceptance Rate) requirements at the higher level, also many technical barriers need to solve.
The content of the invention
The problem of being existed according to prior art, the invention discloses a kind of intelligent vena identification with independent learning ability Method and system, implement global training first first with available sample generation training data before model on-line running, complete The optimization of all parameters such as feature extraction, identification, classification.After model on-line running, office is repeated using the sample newly registered Portion trains, and completes the re-optimization of sorting parameter.Register sample arrival specified quantity or user sets the cycle and model is in dimension The dynamic registration sample accumulated after on-line running is periodically utilized during shield simultaneously, global training is carried out once again, completes feature The re-optimization of all parameters such as extraction, feature recognition, tagsort.Trained by above-mentioned periodically global training with irregularly local Combination, the deep learning model of full convolutional neural networks is really possessed independent learning ability, using dynamic registration sample not Disconnected training, constantly improve identifying algorithm improve authentication precision.Disclosed in other this method in sample image processing procedure The a series of images processing side such as a kind of Normalized Grey Level processing, finger angle correct, finger position correction and interception, train of thought enhancing Method, become apparent from the sample image of acquisition.
Brief description of the drawings
, below will be to embodiment or existing for clearer explanation embodiments of the invention or the technical scheme of prior art There is the required accompanying drawing used in technology description to do one and simply introduce, it should be apparent that, drawings in the following description are only Some embodiments of the present invention, for those of ordinary skill in the art, on the premise of not paying creative work, may be used also To obtain other accompanying drawings according to these accompanying drawings.
Fig. 1 is the structure chart for the intelligent vena identification system that the present invention has independent learning ability；
Fig. 2 is the flow chart for the intelligent vein authentication method that the present invention has independent learning ability；
Fig. 3 is the global training flow chart of the vena identification system of the present invention；
Fig. 4 is the local training flow chart of the vena identification system of the present invention；
Fig. 5 is the vena identification flow chart of the vena identification system of the present invention；
Fig. 6 is the image processing flow figure of the vena identification system of the present invention；
Fig. 7 is that the finger contours of the vena identification system of the present invention identify schematic diagram；
Fig. 8 is finger position calibration and the interception schematic diagram of the vena identification system of the present invention；
Fig. 9 is the finger train of thought enhancing effect schematic diagram of the vena identification system of the present invention；
Figure 10 is the structure chart of the deep learning model of full convolutional neural networks in the present invention；
Figure 11 is the schematic diagram of embodiment 1 in the present invention；
Figure 12 is the schematic diagram of embodiment 2 in the present invention.
Embodiment
To make the purpose, technical scheme and advantage of embodiments of the invention clearer, with reference to the embodiment of the present invention In accompanying drawing, the technical scheme in the embodiment of the present invention is clearly completely described：
A kind of intelligent vein authentication method with independent learning ability as seen in figs. 2-10, specifically includes following step Suddenly：
S1：The vein image information for the multiple samples chosen is gathered, each sample carries out multi collect and carries out sequence number mark Note；
S2：Vein image processing is carried out to the sample information collected：Handled using Normalized Grey Level, the hand of bounds checking Refer to outline identification, the angle calibration system based on finger center line, position correction and interception, the method that the train of thought based on frequency filtering is strengthened Vein image is handled；
S3：To in S2 pass through image procossing formed treat training data carry out left and right angle rotation and peripheral direction translation The key words sorting of data extending and training data；
S4：Training data after mark is inputted to full convolutional neural networks deep learning model and instructed totally according to iteration Practice, carry out global training respectively according to system mode untill the recognition correct rate of model reaches training requirement, then to completing The full convolutional neural networks deep learning model of training carries out local training, and Optimal Parameters caused by training are preserved；
S5：New vein image is entered by full convolutional neural networks deep learning model using the Optimal Parameters described in S4 Row feature extraction, feature recognition and tagsort complete the certification of vein image.
First have to utilize the finger vena sample generation accumulated before full convolutional neural networks deep learning model running Training data carries out once global training to system, to complete the determination of the parameter of feature extraction, feature recognition and classification and optimization. After completing first global training, model can on-line running.After model enters row vein registration, using registering, sample generation is local Training data carries out irregular local training to model, to complete the further optimization of classified part parameter.When model is carried out During vena identification, using global and local train caused by Optimal Parameters, feature extraction is carried out to image to be certified, feature is known Not and classify, realize fast and accurately vena identification.By certain number repeatedly after, the model can accumulate a number of Xindeng Remember sample.When under given conditions for example：After accumulation newly registers sample arrival specified quantity or arrival user specifies the cycle, and And model be in user setting maintenance during when, model using accumulation dynamic registration sample, carry out global training again, it is complete Into the suboptimization again of all parameters such as feature extraction, identification, classification.System passes through the periodically global training when vena identification With the irregular repetition training locally trained and be combined, the wherein parameter of feature extraction, identification and disaggregated model will be continued to optimize, Certification effect also constantly improves, and forms a kind of intelligent authentication system with independent learning ability.
Full convolutional neural networks deep learning model is inputted to the sample collected and carries out once global training, overall situation instruction Practice in the following way：
Gather global training sample.Global training sample M should at least 100, can be from each by collector right-hand man Forefinger, middle finger and nameless collection.Each sample collection 10 times is simultaneously labeled.
Image procossing is carried out to the sample image collected.The global training image of collection by Normalized Grey Level processing, The image procossing such as angle calibration system, position correction and the interception of finger center line, train of thought enhancing, forms the global training data in M × 10.
Training data expands：The each global training data formed through image procossing, then pass through the angle of left and right 1~5 degree 10 After the rotation of degree, then the image enhancement processing to 8 direction translation 1mm around, training data after 99 expansions can be formed.It is global The sample of training is total to can reach M × 990.
Training data key words sorting：990 data carry out classification annotation corresponding to each global training sample, according to random The principle of distribution, 500 data are for training, 490 data are used to train test.Formed M class × 500 training data with M class × 490 test data.
Input after the training data random alignment for completing key words sorting to full convolutional neural networks deep learning model, is entered Row path=1 is totally according to repetitive exercise, untill the recognition correct rate of model reaches training requirement.100 training are often completed, M × 490 test data formed using key words sorting, is checked the recognition correct rate of model.
When checking the accuracy of training pattern in the following way：
Accuracy inspection is identified using cross entropy loss function, convergence is iterated using stochastic gradient descent method, Make cross entropy loss reduction.Because softmax disaggregated models can be defined as formula 1,
Wherein, ak(x) activation for expression in location of pixels x feature passage K.K be classification number, pk(x) it is approximate pole Big value function.I.e. for certain classification k, maximum activation pk(x) ≈ 1, for classification beyond k, akAnd p (x)k(x)≈0.Cross entropy pair Each position pl(x)(x) deviation is formula 2,
Wherein, l:Ω → { 1 ..., K } is the classification of Pixel-level,For the weight matrix set during training.Profit With morphological operations calculation of boundary conditions, weight matrix formula 3 calculates.
Wherein,For classify frequency equilibrium valve weight matrix,Nearest border is represented,Represent secondary near border.W in the present invention0=10, σ ≈ 5.
Loss function J (θ) calculates according to formula 4 in stochastic gradient descent method,
Wherein, θ is parameter, xi、yiFor output, n be training data number,
The corresponding gradient of the loss function of each training sample is calculated by formula 6, and gradient is not after limited number of time iteration It is disconnected to decline, obtain optimized parameter.
Optimal Parameters：Using moving average model, the parameter of full convolutional neural networks deep learning model is optimized, After Model Identification accuracy reaches 99.9%, deconditioning.Shadow parameter (shadow is used in moving average model Variable the parameter of each training to be updated) is corresponded to, is that shadow parameter assigns initial value using formula 7.
Shadow_variable=decay × shadow_variable+ (1-decay) 7
Wherein, decay controls the speed of model modification, more big more tend towards stability.In practice, decay typically can It is arranged to be sufficiently close to 1 constant (0.99 or 0.999).Update faster in the starting stage of training in order that obtaining model, lead to Formula (8) and num_updates parameters are crossed decay size is set dynamically.
Each layer parameter of all full convolutional neural networks deep learning models by optimization is preserved, so as to local training When use.
Model can be enable with whole parameters of the feature extraction of Optimized model, feature recognition and classification by overall situation training Enough carry out the feature extraction of vein image and preliminary feature recognition.In order that the classification that system completes new vein image is known Not, it is also necessary to carry out following local training.
Register sample collection：Need to carry out registration sample collection before vena identification, the system is using registering sample carry out office Train in portion.Registering sample can be from the forefinger of registration of personnel right-hand man, middle finger or nameless different finger collection 2 times.
Image procossing：The registration sample image of collection by Normalized Grey Level processing, finger angle correct, position correction and The image procossings such as interception, train of thought enhancing, form 2 local training datas.
Training data expands：The each local training data formed through image procossing, then pass through 1~5 degree of left and right 10 Angle rotates, then the image enhancement processing to 8 direction translation 1mm around, can form 99 expansion training datas.Local training Data add up to 2 × 99=198.
Training data key words sorting：Key words sorting, according to the principle being randomly assigned, 150 are carried out to local training data Data are for training, 48 data are used to train test.
Import Optimal Parameters：When carrying out local training, the optimization ginseng preserved during global training is imported from storage part first Number.
150 training datas formed using training data key words sorting, input is to full convolutional Neural net after random alignment Network deep learning model carries out path=1 totally according to repetitive exercise, untill the recognition correct rate of model reaches training requirement. 10 training are often completed, 48 test datas formed using training data key words sorting, the discrimination of model are checked.
Optimal Parameters：Using moving average model, the parameter of full convolutional neural networks deep learning model is optimized, After pattern-recognition accuracy reaches 99.9%, deconditioning.
Preserved by the full convolutional neural networks deep learning model parameter of optimization to storage part, to be used during certification.
By local training can further Optimized model sorting parameter, to adapt to dynamically increased classification samples, make Model can carry out the feature extraction and feature recognition and classification of new vein image.
The specific implementation flow of vena identification is as shown in Figure 5
Further, certification sample collection.Need to select arbitrarily from 2 fingers of registration sample collection during vena identification One finger collection certification sample, each sample collection 1 time.
Image procossing：The certification sample image of collection by Normalized Grey Level processing, finger angle correct, position correction and The image procossings such as interception, train of thought enhancing, form authentication data.
Import Optimal Parameters：The optimization ginseng preserved when before vena identification is implemented from storage part importing global and local training Number.
Feature extraction and identification：The optimization that full convolutional neural networks deep learning model is imported using importing in Optimal Parameters Parameter, eigenmatrix is extracted in the authentication data exported from image procossing, eigenmatrix is identified and classification processing, shape Into the class vector of authentication data.
Authentication output result：System confirms to recognize according to the class vector and certification threshold values that are formed in feature extraction and identification Whether card data belong to the sample type of registration, and authentication output by or authentification failure authentication result.
Foregoing training sample, registration sample, certification sample collection are carried out by harvester.Harvester has collection Guiding function, guiding collection finger are put into acquisition zone in the right direction.But this does not ensure that all samples according to identical Angle is acquired in identical position.Simultaneously because the difference of individual, the brightness of image of image pickup part collection can also vary with each individual. In order to eliminate the factor of these influence authentication precisions, it is necessary to carry out image procossing to the vein image of collection.
Image procossing specific implementation flow is as shown in Figure 6：
Picture gray proces；Gray proces are normalized in the vein image that image pickup part gathers using weighted mean method, Obtain the gray-scale map of vein image.
Finger outline identifies:The vein gray level image obtained using picture gray proces, passes through following bounds checking side Method obtains finger contours.It is multiplied using the Sobel Operator of formula 9 and formula 10 with image data matrix A, with the ladder of formula 11 Degree carries out bounds checking.More than pixel of the pixel as finger contours of threshold values, then using linear interpolation method by wire-frame image Vegetarian refreshments connects into line, obtains finger contours.
Finger angle calibration system:After finger contours identification, finger center line is calculated by the following method, as shown in Figure 7.It is true first Determine the coordinate (x of finger tip0, y0), calculate any 2 points of (x on finger center line secondly by formula 12, formula 131, y2) and (x2, y2) coordinate.
Wherein, h is the height of image, and r (y), l (y) are the x coordinate of finger left and right edges corresponding to height y.By above-mentioned Finger center line can carry out the correction of finger angle.Finger position is calibrated and interception.Utilize finger angle in finger angle calibration system Image after correction, according to finger position correction and the interception of authentication region can be carried out shown in Fig. 8.On finger tip and authentication region The distance on side is 100pixels, and the distance on center line and authentication region or so side is 48pixels, and authentication region is highly 272pixels.The authentication region size corrected by finger position is 272 × 96pixels.Even if finger position has deviation, It can ensure that the region of interception is relatively fixed, realize finger position correction for drift.
Train of thought strengthens:Arteries and veins is realized by the following method in the authentication region image intercepted after finger position calibration and intercepting process Network is emphasized, preferably to embody the personal characteristics of sample, improves authentication precision.The gray value that image is often gone first carries out Fu In leaf transformation, then using formula 14 operator carry out frequency filtering, finally by inverse Fourier transform obtain train of thought enhancing Image.As shown in Figure 9
Full convolutional neural networks deep learning model is made up of different types of 35 layer network, and the 1st layer is that input layer can be defeated Enter the gray scale vein image of arbitrary dimension.2nd layer to the 17th layer include 5 groups of two-fold laminations and pond layer, 1 prevent over-fitting Dropout layers, the 18th layer to the 34th layer includes 4 groups of warp laminations, fused layer, 2 layers of warp lamination and 1 single-layer back convolutional layer, Add 4 up-samplings to handle by 4 down-samplings, realize the feature extractions and feature recognition of 16 times of amplifications of original image, output with it is defeated Enter picture size identical eigenmatrix.35th layer is classification layer, realizes tagsort, output and registration number of samples identical Class vector.Concrete structure is as shown in Figure 10：
101input
Input layer, by the greyscale image data of image procossing, each element is 8 gray values of each pixel for input (0-255).Export the matrix of H × W × 1.Wherein H is picture height number of pixels, W is picture width number of pixels, the member of matrix Element is the relative gray values (the double-precision floating number between 0~1) of each pixel.
102conv1-1
First convolutional layer 1, filter size are 3 × 3, depth 64, step-length 1；Activation primitive is used as using ReLU；Square Battle array border is filled using full 0, to keep the size of convolutional layer propagated forward matrix of consequence consistent with current layer matrix size；Using Xavier normal distributions initialize, and parameter is by 0 average, standard deviation) normal distribution produce, wherein fan_ In is the fan-in of weight tensor.Export the matrix of H × W × 64.
103conv1-2
First convolutional layer 2, filter size are 3 × 3, depth 64, step-length 1；Activation primitive is used as using ReLU；Square Filled using full 0 on battle array border；Initialized using Xavier normal distributions.Export the matrix of H × W × 64.
104pool1
First pond layer, using maximum pond method, filter size is 2 × 2, step-length 1, no matrix boundaries filling. Export as the matrix of H/2 × W/2 × 64.
105conv2-1
Second convolutional layer 1, filter size are 3 × 3, depth 128, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/2 × W/2 × 128.
106conv2-2
Second convolutional layer 2, filter size are 3 × 3, depth 128, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/2 × W/2 × 128.
107pool2
Second pond layer, using maximum pond method, filter size is 2 × 2, step-length 1, no matrix boundaries filling. Export as the matrix of H/4 × W/4 × 128.
108conv3-1
3rd convolutional layer 1, filter size are 3 × 3, depth 256, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/4 × W/4 × 256.
109conv3-2
3rd convolutional layer 2, filter size are 3 × 3, depth 256, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/4 × W/4 × 256.
1010pool3
3rd pond layer, using maximum pond method, filter size is 2 × 2, step-length 1, no matrix boundaries filling. Export as the matrix of H/8 × W/8 × 256.
1011conv4-1
Volume Four lamination 1, filter size are 3 × 3, depth 512, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/8 × W/8 × 512.
1012conv4-2
Volume Four lamination 2, filter size are 3 × 3, depth 512, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/8 × W/8 × 512.
1013dropout4
Dropout layers, apply 50% Dropout to input data, prevent over-fitting.
1014pool4
4th pond layer, using maximum pond method, filter size is 2 × 2, step-length 1, no matrix boundaries filling. Export as the matrix of H/16 × W/16 × 512.
1015conv5-1
5th convolutional layer 1, filter size are 3 × 3, depth 1024, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/16 × W/16 × 1024.
1016conv5-2
5th convolutional layer 2, filter size are 3 × 3, depth 1024, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/16 × W/16 × 1024.
1017dropout5
Dropout layers, apply 50% Dropout to input data, prevent over-fitting.
1018conv6-1
First warp lamination 1, filter size are 2 × 2, depth 512, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.2 times of up-samplings are carried out simultaneously (upsampling) matrix of H/8 × W/8 × 512, is exported.
1019merge6
Fused layer, the output after the output and 1013 Dropout processing of 1018 the first warp lamination 1 is merged, along Last dimension is spliced, the output matrix of H/8 × W/8 × 1024.
1020conv6-2
First warp lamination 2, filter size are 3 × 3, depth 512, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/8 × W/8 × 512.
1021conv6-3
First warp lamination 3, filter size are 3 × 3, depth 512, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/8 × W/8 × 512.
1022conv7-1
Second warp lamination 1, filter size are 2 × 2, depth 256, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.2 times of up-samplings are carried out simultaneously (upsampling) matrix of H/4 × W/4 × 256, is exported.
1023merge7
Fused layer, the output of 109 the 3rd convolutional layer 2 and the output of 1022 the second warp lamination 1 are merged, along last One dimension is spliced, the output matrix of H/4 × W/4 × 512.
1024conv7-2
Second warp lamination 2, filter size are 3 × 3, depth 256, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/4 × W/4 × 256.
1025conv7-3
Second warp lamination 3, filter size are 3 × 3, depth 256, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/4 × W/4 × 256.
1026conv8-1
3rd warp lamination 1, filter size are 2 × 2, depth 128, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.2 times of up-samplings are carried out simultaneously (upsampling) matrix of H/2 × W/2 × 128, is exported.
1027merge8
Fused layer, 106 output of the second convolutional layer 2 and the output of 1026 the 3rd warp laminations 1 are merged, along last Individual dimension is spliced, the output matrix of H/2 × W/2 × 256.
1028conv8-2
3rd warp lamination 2, filter size are 3 × 3, depth 128, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/2 × W/2 × 128.
1029conv8-3
3rd warp lamination 3, filter size are 3 × 3, depth 128, step-length 1；Activation letter is used as using ReLU Number；Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H/2 × W/2 × 128.
1030conv9-1
4th warp lamination 1, filter size are 2 × 2, depth 64, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.2 times of up-samplings (upsampling) are carried out simultaneously, it is defeated Go out the matrix of H × W × 64.
1031merge9
Fused layer, the output of 103 the first convolutional layer 2 and the output of 1030 the 4th warp lamination 1 are merged, along last One dimension is spliced, the output matrix of H × W × 128.
1032conv9-2
4th warp lamination 2, filter size are 3 × 3, depth 64, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H × W × 64.
1033conv9-3
4th warp lamination 3, filter size are 3 × 3, depth 64, step-length 1；Activation primitive is used as using ReLU； Matrix boundaries are filled using full 0；Initialized using Xavier normal distributions.Export the matrix of H × W × 64.
1034conv10-1
5th warp lamination, filter size are 3 × 3, depth 1, step-length 1；Activation primitive is used as using ReLU；Square Filled using full 0 on battle array border；Initialized using Xavier normal distributions.Export the feature matrix of H × W × 1.
1035conv10-2
Classification layer, filter size is that 1 × 1, depth is M (logging in number of samples), step-length 1；Using softmax as Activation primitive；Matrix boundaries are without filling.Export and tie up identification vector for M, represent identified sample and log in the matching probability of sample.
A kind of intelligent vena identification system with independent learning ability as shown in Figure 1, including：
Gather the image unit of vein image information and the light source of brightness is provided in gatherer process；
The graphics processing unit of the image information of the image unit transmission is received, described image processing unit is to receiving Image using Normalized Grey Level processing, the finger contours identification of bounds checking, the angle calibration system based on finger center line, position school Just and the method strengthened of interception, train of thought based on frequency filtering is handled；
Receive the image information of described image processing unit transmission, input the image information received to full convolutional Neural The deep learning model of network is iterated the recognition unit of training, and the recognition unit is to full convolutional neural networks deep learning Model carries out global training untill the recognition correct rate of model reaches training requirement respectively, then to completing the full convolution of training Neutral net deep learning model carries out local training, and Optimal Parameters caused by training are preserved, using after training model it is excellent Change parameter, carry out feature extraction, feature recognition and the classification of vena identification image, complete vena identification；
Receive the Optimal Parameters information of recognition unit transmission and the authentication image information of graphics processing unit transmission Memory cell.
Further, the recognition unit is when the deep learning model to full convolutional neural networks is trained：Using The finger vena sample generation training data accumulated carries out once global training to system, completes feature extraction, feature recognition Determine and optimize with the parameter of tagsort, after completing first global training, model can on-line running:Stepped on when model enters row vein After note, the local training irregular to model progress of local training data is generated using sample is registered, completes classified part parameter Further optimization；
The recognition unit to vein image when being authenticated：Vein image after image procossing is inputted to complete The deep learning model of convolutional neural networks is carried out using Optimal Parameters caused by global and local training to image to be certified Feature extraction, feature recognition and classification.
By the intelligent vena identification system disclosed by the invention with independent learning ability to full convolutional neural networks Deep learning model carry out global training and local training make the deep learning model of full convolutional neural networks really possess from Primary learning ability, and feature extraction and the tagsort parameter of optimization are preserved, led when carrying out vena identification from memory cell The Optimal Parameters that preserve are authenticated when entering global and local training, authentication output by or authentification failure authentication result.
System realization scheme 1：
A kind of intelligent vena identification system with independent learning ability is as shown in figure 11, including harvester, leads to Believe interface and server composition, wherein harvester passes through communication interface and server real-time data communication.Wherein harvester Include light source and image unit, graphics processing unit, recognition unit and memory cell are wherein provided with server.
Harvester completes the collection of finger venous image, and the view data of collection is sent into clothes by communication interface The graphics processing unit of business device.
Light source is made up of the near-infrared LED that 7 wavelength are 890nm, is controlled luminous by recognition unit or is extinguished.
Image unit is made up of wide-angle lens, infrared ray filter plate and high sensitivity infrared photography CMOS, and data-interface follows USB2.0 standards, such as Sony XC-E150 video cameras.Recognition unit control image unit collection vein image, after gathering image Data are sent to the graphics processing unit of server by data-interface, communication interface.Gather the size of image for 640 × 480pixels。
Communication interface can be wired serial communication, such as USB, LAN, RS232 or radio communication, as WIFI, Bluetooth, Zigbee etc..Communications interface transmission server recognition unit is sent to the control data of light source and image unit, and shooting list Member is sent to the view data of graphics processing unit.
Server is by two-way to the strong V3 CPU of E5 2683,128G DDR4 internal memories, Intel C612 chipsets mainboard, 1T Solid state hard disc, 4 road NVIDIA GTX TITAN XP video cards (GPU) composition, can be achieved 28 kernels, 56 threads concurrent operation. Including foregoing graphics processing unit, recognition unit and memory cell.Complete foregoing image procossing, full convolutional neural networks Overall situation training, local training and certification, storage optimization parameter, the vein image and training data of collection.
The authentication region size H × W intercepted after image procossing is 272 × 96pixels.Global training sample is derived from 25 people Right-hand man's forefinger and middle finger, form 100 samples (species), each sample collection 10 times, form 1000 training datas, then Handled by training data expansion, form 99000 training datas.Wherein 50000 data are used for the global training of model, 49000 data are used for the test of global training.Learning rate is set to 0.0001, and recognition correct rate terminates the overall situation when reaching 99.9% Training.
Local training uses 2 data for newly registering sample, then is handled by training data expansion, forms 198 training Data.Wherein 150 data are used for model and locally trained, and 49 data are used for the test locally trained.Learning rate is set to 0.0001, recognition correct rate terminates global training when reaching 99.9%.Threshold values during identification is set to 0.0001.
When the training data for newly registering sample has reached 100, and in system maintenance time section (such as AM1:00), it is System carries out foregoing global training again.
Collecting device, which can be distributed, in the implementation is arranged on different zones, and takes full advantage of the powerful computing of server Function and mass storage function, it is adapted to large-scale distributed application.Harvester cost simple in construction is relatively low simultaneously, is adapted to big rule Mould popularization and application.
System realization scheme 2：
System realization scheme 2 is as shown in figure 12, is made up of authentication device, communication interface and server.Wherein authentication device Formed including light source, image unit, CPU/ processors and memory cell.
Light source is made up of the near-infrared LED that 7 wavelength are 890nm, is controlled luminous by identification part 1 or is extinguished.
Image unit is made up of wide-angle lens, infrared ray filter plate and high sensitivity infrared photography CMOS, and data-interface follows USB2.0 standards, such as Sony XC-E150 video cameras.Recognition unit control image unit collection vein image, after gathering image Data are sent to the graphics processing unit of CPU/ processors by data-interface.Gather the size of image for 640 × 480pixels。
CPU/ processors can be the processor that MCU, DSP, FPGA of embedded system etc. have computing and control function. Employ the AM3358 dsp chips of TI companies in the implementation case, integrated chip ARM Cortex-A8MCU kernels, tool There is 3D to accelerate processing function.Built-in 176KB ROM and 64KB RAM, external 1GB DDR3RAM, 4MB EEPROM and 64GB SD Storage card.With USB2.0, WIFI, 100,000,000 LAN, bluetooth communication interface.Include foregoing graphics processing unit and identification is single Member.Image processing part completes foregoing a series of images processing, and recognition unit completes vein using foregoing full convolutional neural networks Feature extraction, feature recognition and classification during certification.The full convolutional neural networks of recognition unit use shape after server end training Into Optimal Parameters.
Memory cell can be in the piece of embedded system or the storage device such as piece outer RAM, erasable ROM, SD card, uses Come Optimal Parameters that storage server sends and the image data of authentication device collection.
Communication interface can be wired serial communication, such as USB, LAN, RS232 or radio communication, as WIFI, Bluetooth, Zigbee etc..Communications interface transmission server is sent to the Optimal Parameters of authentication device, and authentication device is sent to server View data.
Server forms identical with system realization scheme 1 with function.
By authentication device gather vein image, the authentication region size H × W intercepted after image procossing be 272 × 96pixels.Global training sample is derived from the right-hand man's forefinger and middle finger of 25 people, forms 100 samples (species), each sample Collection 10 times, 1000 training datas are formed, then handled by training data expansion, form 99000 training datas.Wherein 50000 data are used for the global training of model, and 49000 data are used for the test of global training.Overall situation training is complete by server Into learning rate is set to 0.0001, and recognition correct rate terminates global training when reaching 99.9%.
Local training uses 2 data for newly registering sample, then is handled by training data expansion, forms 198 training Data.Wherein 150 data are used for model and locally trained, and 49 data are used for the test locally trained.Learning rate is set to 0.0001, recognition correct rate terminates global training when reaching 99.9%.Threshold values during identification is set to 0.0001.
When the training data for newly registering sample has reached 100, and in system maintenance time section (such as AM1:00), it is System carries out foregoing global training again.
The implementation will gather and authentication function module is integrated in an embedded device, has small volume, carries Easy for installation, environment adapts to the features such as strong.Model training and data storage are placed in server, realize the backstage of various training Implementation and the concentration magnanimity management of data.Both it had been adapted to unit work, and had been adapted to large scale network application again.
A series of images treatment technology is employed in the present invention, gray proces, finger angle is normalized in authentication image Degree correction, finger position correction and train of thought emphasize that the noise for eliminating effect characteristicses extraction and feature recognition to greatest extent is done Disturb and retain the characteristic information of artwork, make vein personal characteristics more obvious, be effectively improved feature extraction and feature recognition Precision.The deep learning of the full convolutional neural networks of the feature extraction that can be accurate to pixel scale and feature recognition is employed again Model, further increase the vena identification degree of accuracy.Its certification error rate EER is reduced to less than 0.1% from 0.39%, while bright It is aobvious to reduce false acceptance rate.The vena identification system false acceptance rate of the present invention can be controlled below 0.001%, reach essence The level of level of confidentiality authentication.
The foregoing is only a preferred embodiment of the present invention, but protection scope of the present invention be not limited thereto, Any one skilled in the art the invention discloses technical scope in, technique according to the invention scheme and its Inventive concept is subject to equivalent substitution or change, should all be included within the scope of the present invention.

Claims (9)

  1. A kind of 1. intelligent vein authentication method with independent learning ability, it is characterised in that：Comprise the following steps：
    S1：The vein image information for the multiple samples chosen is gathered, each sample carries out multi collect and carries out sequence number mark；
    S2：Vein image processing is carried out to the sample information collected：Handled using Normalized Grey Level, the finger wheel of bounds checking The method that wide identification, the angle calibration system based on finger center line, position correction and interception, the train of thought based on frequency filtering are strengthened is to quiet Arteries and veins image is handled；
    S3：The data for treating training data progress left and right angle rotation and peripheral direction translation formed to passing through image procossing in S2 Expand the key words sorting with training data；
    S4：Training data after mark is inputted to full convolutional neural networks deep learning model and carried out totally according to repetitive exercise, Global training is carried out respectively according to system mode untill the recognition correct rate of model reaches training requirement, then to completing to train Full convolutional neural networks deep learning model carry out local training, Optimal Parameters caused by training are preserved；
    S5：Spy is carried out to new vein image by full convolutional neural networks deep learning model using the Optimal Parameters described in S4 Sign extraction, feature recognition and tagsort complete the certification of vein image.
  2. 2. a kind of intelligent vein authentication method with independent learning ability according to claim 1, is further characterized in that： Be repeated, regularly global training and indefinite is employed when being trained to full convolutional neural networks deep learning model The autonomous learning method that phase, locally training combined：Before full convolutional neural networks deep learning model on-line running, utilization has been accumulated Tired sample carries out once global training, and preserves feature extraction, feature recognition and the tagsort parameter of optimization；Model is reached the standard grade After operation, the irregular local optimization trained, complete tagsort parameter is carried out to the model using the sample newly registered； Under specified conditions, model is carried out using the sample newly accumulated it is regularly global train, again to feature extraction, feature recognition and Tagsort parameter optimizes, and makes vena identification precision with system operation using constantly carrying by above-mentioned autonomous learning method Height, realize intelligent vena identification.
  3. 3. a kind of intelligent vein authentication method with independent learning ability according to claim 1, is further characterized in that： The full convolutional neural networks deep learning model includes different types of 35 layer network, and the 1st layer is to allow to input arbitrary dimension The input layer of gray scale vein image；2nd layer to the 17th layer include 5 groups of two-fold laminations and pond layer, 1 prevent over-fitting Dropout layers, the 18th layer to the 34th layer includes 4 groups of warp laminations, fused layer, 2 layers of warp lamination and 1 single-layer back convolutional layer, Add 4 up-samplings to handle by 4 down-samplings, realize the feature extractions and feature recognition of 16 times of amplifications of original image, output with it is defeated Enter picture size identical eigenmatrix；35th layer is classification layer, realizes tagsort, output and registration number of samples identical Class vector.
  4. 4. a kind of intelligent vein authentication method with independent learning ability according to claim 1, is further characterized in that： Under specified conditions before model on-line running or after on-line running, the sample image collected is inputted to full convolutional Neural net Network deep learning model carries out once global training：Optimize whole parameters of feature extraction, feature recognition and tagsort：
    Gather the vein image information of multiple samples and be labeled；
    Vein image processing is carried out to the sample image after mark；
    Left and right angle rotation is carried out to sample image and the expansion of training image is realized in peripheral direction translation, then to the sample after expansion This image carries out classification annotation, marks the data for training and testing；
    Sample image after mark is inputted to full convolutional neural networks deep learning model to the repetitive exercise for carrying out total evidence, adopted The parameter of full convolutional neural networks deep learning model is optimized with moving average model, until the recognition correct rate of model Untill reaching training requirement, the recognition correct rate of model is checked using cross entropy loss function, using under stochastic gradient Drop method is iterated convergence, and each layer parameter of all full convolutional neural networks deep learning models is preserved.
  5. 5. a kind of intelligent vein authentication method with independent learning ability according to claim 1, is further characterized in that： After model on-line running, irregular office is carried out to full convolutional neural networks deep learning model using the sample image newly registered Portion's training, optimization tagsort parameter, specifically in the following way：
    The vein sample image collected is registered, carrying out vein image processing to above-mentioned sample image forms local training data；
    Data extending is carried out to the vein image of above-mentioned processing：The image enhaucament shape of left rotation and right rotation and peripheral direction is carried out to image Into multiple expansion training datas；Key words sorting is carried out to expanding training data：Mark the data for training and testing；
    Import Optimal Parameters：The Optimal Parameters that preserve import full convolutional neural networks deep learning model and right when the overall situation is trained Treat that training data carries out local training：The parameter of full convolutional neural networks deep learning model is carried out using moving average model Optimization, until the accuracy of model reaches training requirement then deconditioning；
    The model parameter for completing locally to train is preserved.
  6. 6. a kind of intelligent vein authentication method with independent learning ability according to claim 1, is further characterized in that： The certification of the vein image is in the following way：
    The collection of certification sample：Need to select any one finger collection to recognize from the finger of registration sample collection during vena identification Demonstrate,prove sample image；
    Image procossing：Vein image processing formation authentication data is carried out to the certification sample image of collection；
    Import Optimal Parameters：Global training is being imported into complete roll up with the Optimal Parameters preserved during local train before vena identification is implemented Product neutral net deep learning model；
    Feature extraction and identification：Eigenmatrix is extracted in the authentication data exported in extraction image processing process, to eigenmatrix It is identified and classification is handled, forms the class vector of authentication data；
    Authentication output result：Full convolutional neural networks deep learning model is according to the classification formed in feature extraction and identification process Vector sum certification threshold values confirm authentication data whether belong to the sample type of registration and authentication output by or authentification failure recognize Demonstrate,prove result.
  7. 7. a kind of intelligent vein authentication method with independent learning ability according to claim 1, is further characterized in that： The finger contours identification of the bounds checking is in the following way：
    <mrow> <msub> <mi>G</mi> <mi>x</mi> </msub> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>-</mo> <mn>2</mn> </mrow> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>2</mn> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> <mo>&amp;times;</mo> <mi>A</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>9</mn> <mo>)</mo> </mrow> </mrow>
    <mrow> <msub> <mi>G</mi> <mi>y</mi> </msub> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> <mtd> <mrow> <mo>-</mo> <mn>2</mn> </mrow> </mtd> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>0</mn> </mtd> </mtr> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mn>2</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> <mo>&amp;times;</mo> <mi>A</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>10</mn> <mo>)</mo> </mrow> </mrow>
    <mrow> <mi>G</mi> <mo>=</mo> <msqrt> <mrow> <msubsup> <mi>G</mi> <mi>x</mi> <mn>2</mn> </msubsup> <mo>+</mo> <msubsup> <mi>G</mi> <mi>y</mi> <mn>2</mn> </msubsup> </mrow> </msqrt> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>11</mn> <mo>)</mo> </mrow> </mrow>
    It is multiplied using the Sobel Operator of above-mentioned formula (9) and formula (10) with image data matrix A, with the gradient of formula (11) Bounds checking is carried out, more than pixel of the pixel as finger contours of threshold values, then using linear interpolation method by contour pixel Point connects into line, obtains finger contours；
    Angle calibration system based on finger center line, position correction and the intercept method are in the following way：
    <mrow> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>x</mi> <mn>1</mn> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>r</mi> <mn>1</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mn>1</mn> </msub> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>l</mi> <mn>1</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mn>1</mn> </msub> <mo>)</mo> </mrow> </mrow> <mn>2</mn> </mfrac> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>y</mi> <mn>1</mn> </msub> <mo>=</mo> <mfrac> <mrow> <mn>2</mn> <msub> <mi>y</mi> <mn>0</mn> </msub> <mo>+</mo> <mi>h</mi> </mrow> <mn>3</mn> </mfrac> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>12</mn> <mo>)</mo> </mrow> </mrow>
    <mrow> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>x</mi> <mn>2</mn> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>r</mi> <mn>2</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mn>2</mn> </msub> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>l</mi> <mn>2</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mn>2</mn> </msub> <mo>)</mo> </mrow> </mrow> <mn>2</mn> </mfrac> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>y</mi> <mn>2</mn> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>y</mi> <mn>0</mn> </msub> <mo>+</mo> <mn>2</mn> <mi>h</mi> </mrow> <mn>3</mn> </mfrac> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>13</mn> <mo>)</mo> </mrow> </mrow>
    Calculate finger center line：Coordinate (the x of finger tip is determined first0, y0), calculate finger secondly by formula (12), formula (13) Any two points (x on center line1, y2) and (x2, y2) coordinate；Wherein, h is the height of image, and r (y), l (y) are corresponding to height y The x coordinate of finger left and right edges；The correction of finger angle can be carried out by above-mentioned finger center line, using in finger angle calibration system Image after finger angle correct carries out finger position correction and the interception of authentication region, realizes finger position correction for drift；
    The train of thought Enhancement Method based on frequency filtering is in the following way：
    <mrow> <mi>k</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mrow> <mo>(</mo> <mi>i</mi> <mo>&amp;le;</mo> <mi>m</mi> <mo>)</mo> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mrow> <mo>(</mo> <mi>i</mi> <mo>&gt;</mo> <mi>m</mi> <mo>)</mo> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>14</mn> <mo>)</mo> </mrow> </mrow>
    The gray value that image is often gone carries out Fourier transformation, then carries out frequency filtering using the operator of formula (14), finally The image of train of thought enhancing is obtained by inverse Fourier transform.
  8. A kind of 8. intelligent vena identification system with independent learning ability, it is characterised in that including：
    Gather the image unit of vein image information and the light source of brightness is provided in gatherer process；
    Receive the graphics processing unit of the image information of image unit transmission, described image processing unit is to the figure that receives As using Normalized Grey Level processing, the finger contours identification of bounds checking, the angle calibration system based on finger center line, position correction and The method that interception, the train of thought based on frequency filtering are strengthened is handled；
    Receive the image information of described image processing unit transmission, input the image information received to full convolutional neural networks Deep learning model be iterated the recognition unit of training, the recognition unit is to full convolutional neural networks deep learning model Global training is carried out respectively untill the recognition correct rate of model reaches training requirement, then to completing the full convolutional Neural of training Network deep learning model carries out local training, and Optimal Parameters caused by training are preserved, and is joined using the optimization of model after training Number, feature extraction, feature recognition and the classification of vena identification image are carried out, complete vena identification；
    Receive the Optimal Parameters information of the recognition unit transmission and the authentication image information storage list of graphics processing unit transmission Member.
  9. 9. a kind of intelligent vena identification system with independent learning ability according to claim 8, is further characterized in that： The recognition unit is when the deep learning model to full convolutional neural networks is trained：Using the finger vena sample accumulated This generation training data carries out once global training to system, and the parameter for completing feature extraction, feature recognition and tagsort is true Fixed and optimization, after completing first global training, model can on-line running:After model enters row vein registration, using registering sample Generate local training data and the irregular local further optimization trained, complete classified part parameter is carried out to model；It is described Recognition unit to vein image when being authenticated：Vein image after image procossing is inputted to full convolutional neural networks Deep learning model using Optimal Parameters caused by global and local training, feature extraction, spy are carried out to image to be certified Sign identification and classification.
CN201711013178.XA 2017-10-26 2017-10-26 Intelligent vein authentication method and system with autonomous learning capability Active CN107832684B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711013178.XA CN107832684B (en) 2017-10-26 2017-10-26 Intelligent vein authentication method and system with autonomous learning capability

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711013178.XA CN107832684B (en) 2017-10-26 2017-10-26 Intelligent vein authentication method and system with autonomous learning capability

Publications (2)

Publication Number Publication Date
CN107832684A true CN107832684A (en) 2018-03-23
CN107832684B CN107832684B (en) 2021-08-03

Family

ID=61649404

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711013178.XA Active CN107832684B (en) 2017-10-26 2017-10-26 Intelligent vein authentication method and system with autonomous learning capability

Country Status (1)

Country Link
CN (1) CN107832684B (en)

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108898559A (en) * 2018-06-20 2018-11-27 中国科学院光电技术研究所 Atmospheric dispersion modification method based on image deconvolution
CN109034016A (en) * 2018-07-12 2018-12-18 辽宁工业大学 A kind of hand back vein image-recognizing method based on S-CNN model of universality
CN109410143A (en) * 2018-10-31 2019-03-01 泰康保险集团股份有限公司 Image enchancing method, device, electronic equipment and computer-readable medium
CN109472784A (en) * 2018-10-31 2019-03-15 安徽医学高等专科学校 Based on the recognition methods for cascading full convolutional network pathological image mitotic cell
CN109620198A (en) * 2019-02-21 2019-04-16 天津惊帆科技有限公司 Cardiovascular index detection and model training method and device
CN109784376A (en) * 2018-12-26 2019-05-21 中国科学院长春光学精密机械与物理研究所 A kind of Classifying Method in Remote Sensing Image and categorizing system
CN109815869A (en) * 2019-01-16 2019-05-28 浙江理工大学 A kind of finger vein identification method based on the full convolutional network of FCN
CN109871859A (en) * 2018-09-28 2019-06-11 北京矩视智能科技有限公司 One kind automatically generating training set of images system
CN110414373A (en) * 2019-07-08 2019-11-05 武汉大学 A kind of deep learning palm vein identification system and method based on cloud side end cooperated computing
CN110543822A (en) * 2019-07-29 2019-12-06 浙江理工大学 finger vein identification method based on convolutional neural network and supervised discrete hash algorithm
CN111639559A (en) * 2020-05-15 2020-09-08 圣点世纪科技股份有限公司 Finger vein machine learning identification method and device based on topographic concave-convex characteristics
CN111967351A (en) * 2020-07-31 2020-11-20 华南理工大学 Deep tree network-based finger vein authentication algorithm, device, medium and equipment
JP2021533493A (en) * 2019-04-03 2021-12-02 平安科技(深▲せん▼)有限公司Ping An Technology (Shenzhen) Co., Ltd. Finger vein matching method, device, computer equipment, and storage medium
CN113805480A (en) * 2020-06-11 2021-12-17 丰田自动车株式会社 Machine learning device, machine learning system, machine learning method, and storage medium

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101719222A (en) * 2009-11-27 2010-06-02 北京中星微电子有限公司 Method and device for training classifiers and method and device for identifying human face
CN102222229A (en) * 2011-07-28 2011-10-19 陈庆武 Method for preprocessing finger vein images
CN104063645A (en) * 2014-07-01 2014-09-24 清华大学深圳研究生院 Identity recognition method based on ECG dynamic self-updating samples
CN105137389A (en) * 2015-09-02 2015-12-09 安宁 Video-assisted radiofrequency positioning method and apparatus
CN105975974A (en) * 2016-05-10 2016-09-28 深圳市金脉智能识别科技有限公司 ROI image extraction method in finger vein identification
US9530052B1 (en) * 2013-03-13 2016-12-27 University Of Maryland System and method for sensor adaptation in iris biometrics
CN106326886A (en) * 2016-11-07 2017-01-11 重庆工商大学 Finger-vein image quality evaluation method and system based on convolutional neural network
CN106529468A (en) * 2016-11-07 2017-03-22 重庆工商大学 Finger vein identification method and system based on convolutional neural network
CN106874581A (en) * 2016-12-30 2017-06-20 浙江大学 A kind of energy consumption of air conditioning system in buildings Forecasting Methodology based on BP neural network model
CN106971174A (en) * 2017-04-24 2017-07-21 华南理工大学 A kind of CNN models, CNN training methods and the vein identification method based on CNN
CN106991368A (en) * 2017-02-20 2017-07-28 北京大学 A kind of finger vein checking personal identification method based on depth convolutional neural networks

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101719222A (en) * 2009-11-27 2010-06-02 北京中星微电子有限公司 Method and device for training classifiers and method and device for identifying human face
CN102222229A (en) * 2011-07-28 2011-10-19 陈庆武 Method for preprocessing finger vein images
US9530052B1 (en) * 2013-03-13 2016-12-27 University Of Maryland System and method for sensor adaptation in iris biometrics
CN104063645A (en) * 2014-07-01 2014-09-24 清华大学深圳研究生院 Identity recognition method based on ECG dynamic self-updating samples
CN105137389A (en) * 2015-09-02 2015-12-09 安宁 Video-assisted radiofrequency positioning method and apparatus
CN105975974A (en) * 2016-05-10 2016-09-28 深圳市金脉智能识别科技有限公司 ROI image extraction method in finger vein identification
CN106326886A (en) * 2016-11-07 2017-01-11 重庆工商大学 Finger-vein image quality evaluation method and system based on convolutional neural network
CN106529468A (en) * 2016-11-07 2017-03-22 重庆工商大学 Finger vein identification method and system based on convolutional neural network
CN106874581A (en) * 2016-12-30 2017-06-20 浙江大学 A kind of energy consumption of air conditioning system in buildings Forecasting Methodology based on BP neural network model
CN106991368A (en) * 2017-02-20 2017-07-28 北京大学 A kind of finger vein checking personal identification method based on depth convolutional neural networks
CN106971174A (en) * 2017-04-24 2017-07-21 华南理工大学 A kind of CNN models, CNN training methods and the vein identification method based on CNN

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
HUAFENG QIN: "Deep representation-based feature extraction and recovering for finger-vein verification", 《JOURNAL OF LATEX CLASS FILES》 *
庞晓红: "指静脉身份识别算法的研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 *

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108898559B (en) * 2018-06-20 2021-11-19 中国科学院光电技术研究所 Atmospheric dispersion correction method based on image deconvolution
CN108898559A (en) * 2018-06-20 2018-11-27 中国科学院光电技术研究所 Atmospheric dispersion modification method based on image deconvolution
CN109034016A (en) * 2018-07-12 2018-12-18 辽宁工业大学 A kind of hand back vein image-recognizing method based on S-CNN model of universality
CN109034016B (en) * 2018-07-12 2021-10-15 辽宁工业大学 Universal S-CNN model-based hand back vein image identification method
CN109871859A (en) * 2018-09-28 2019-06-11 北京矩视智能科技有限公司 One kind automatically generating training set of images system
CN109410143A (en) * 2018-10-31 2019-03-01 泰康保险集团股份有限公司 Image enchancing method, device, electronic equipment and computer-readable medium
CN109472784A (en) * 2018-10-31 2019-03-15 安徽医学高等专科学校 Based on the recognition methods for cascading full convolutional network pathological image mitotic cell
CN109784376A (en) * 2018-12-26 2019-05-21 中国科学院长春光学精密机械与物理研究所 A kind of Classifying Method in Remote Sensing Image and categorizing system
CN109815869A (en) * 2019-01-16 2019-05-28 浙江理工大学 A kind of finger vein identification method based on the full convolutional network of FCN
CN109620198A (en) * 2019-02-21 2019-04-16 天津惊帆科技有限公司 Cardiovascular index detection and model training method and device
JP7141518B2 (en) 2019-04-03 2022-09-22 平安科技（深▲せん▼）有限公司 Finger vein matching method, device, computer equipment, and storage medium
US11893773B2 (en) 2019-04-03 2024-02-06 Ping An Technology (Shenzhen) Co., Ltd. Finger vein comparison method, computer equipment, and storage medium
JP2021533493A (en) * 2019-04-03 2021-12-02 平安科技(深▲せん▼)有限公司Ping An Technology (Shenzhen) Co., Ltd. Finger vein matching method, device, computer equipment, and storage medium
CN110414373B (en) * 2019-07-08 2021-09-24 武汉大学 Deep learning palm vein recognition system and method based on cloud edge-side cooperative computing
CN110414373A (en) * 2019-07-08 2019-11-05 武汉大学 A kind of deep learning palm vein identification system and method based on cloud side end cooperated computing
CN110543822A (en) * 2019-07-29 2019-12-06 浙江理工大学 finger vein identification method based on convolutional neural network and supervised discrete hash algorithm
CN111639559A (en) * 2020-05-15 2020-09-08 圣点世纪科技股份有限公司 Finger vein machine learning identification method and device based on topographic concave-convex characteristics
WO2021227331A1 (en) * 2020-05-15 2021-11-18 圣点世纪科技股份有限公司 Finger vein machine learning recognition method and apparatus based on terrain concave and convex chjaracteristic
CN111639559B (en) * 2020-05-15 2022-07-05 圣点世纪科技股份有限公司 Finger vein machine learning identification method and device based on topographic concave-convex characteristics
CN113805480A (en) * 2020-06-11 2021-12-17 丰田自动车株式会社 Machine learning device, machine learning system, machine learning method, and storage medium
CN111967351B (en) * 2020-07-31 2023-06-20 华南理工大学 Finger vein authentication algorithm, device, medium and equipment based on depth tree network
CN111967351A (en) * 2020-07-31 2020-11-20 华南理工大学 Deep tree network-based finger vein authentication algorithm, device, medium and equipment

Also Published As

Publication number Publication date
CN107832684B (en) 2021-08-03

Similar Documents

Publication Publication Date Title
CN107832684A (en) A kind of intelligent vein authentication method and system with independent learning ability
CN109543606B (en) Human face recognition method with attention mechanism
CN107016370B (en) A kind of partial occlusion face identification method based on data enhancing
Yuan et al. Fingerprint liveness detection using an improved CNN with image scale equalization
CN106548159A (en) Reticulate pattern facial image recognition method and device based on full convolutional neural networks
CN106529468A (en) Finger vein identification method and system based on convolutional neural network
WO2019179403A1 (en) Fraud transaction detection method based on sequence width depth learning
CN106991368A (en) A kind of finger vein checking personal identification method based on depth convolutional neural networks
CN108235770A (en) image identification method and cloud system
CN106096535A (en) A kind of face verification method based on bilinearity associating CNN
CN108090830B (en) Credit risk rating method and device based on facial portrait
CN110111113B (en) Abnormal transaction node detection method and device
CN106709450A (en) Recognition method and system for fingerprint images
CN106408039A (en) Off-line handwritten Chinese character recognition method carrying out data expansion based on deformation method
CN108564040B (en) Fingerprint activity detection method based on deep convolution characteristics
CN106651373A (en) Method and device for establishing mixed fraudulent trading detection classifier
CN105389486B (en) A kind of authentication method based on mouse behavior
CN106875007A (en) End-to-end deep neural network is remembered based on convolution shot and long term for voice fraud detection
CN109033945A (en) A kind of human body contour outline extracting method based on deep learning
CN107169479A (en) Intelligent mobile equipment sensitive data means of defence based on fingerprint authentication
CN102254188A (en) Palmprint recognizing method and device
CN102270308A (en) Facial feature location method based on five sense organs related AAM (Active Appearance Model)
CN109977832A (en) A kind of image processing method, device and storage medium
CN111260568B (en) Peak binarization background noise removing method based on multi-discriminator countermeasure network
CN106709500A (en) Image feature matching method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant