CN107563294A - A kind of finger vena characteristic extracting method and system based on self study - Google Patents
A kind of finger vena characteristic extracting method and system based on self study Download PDFInfo
- Publication number
- CN107563294A CN107563294A CN201710656587.5A CN201710656587A CN107563294A CN 107563294 A CN107563294 A CN 107563294A CN 201710656587 A CN201710656587 A CN 201710656587A CN 107563294 A CN107563294 A CN 107563294A
- Authority
- CN
- China
- Prior art keywords
- layer
- finger vein
- identified
- finger
- coding
- 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.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of finger vena characteristic extracting method and system based on self study, wherein, the finger vena characteristic extracting method includes：The finger vein image to be identified got is pre-processed, obtains pretreated finger vein image to be identified；Feature extraction processing is carried out to the pretreated finger vein image to be identified using own coding neural network model, obtains finger vein pattern to be identified.In embodiments of the present invention, feature extraction is carried out by the finger vein feature extraction method based on self study, can be in the case where ensuring that finger hand vein recognition precision is enough, learnt automatically by machine to the finger vein pattern with descriptive and high discrimination, work difficulty is reduced, improves matching precision.
Description
Technical field
The present invention relates to biometrics identification technology field, more particularly to a kind of finger vein pattern extraction based on self study
Method and system.
Background technology
It is feature inside people to refer to vein, because its high security and convenience make it that referring to vein identification technology turns into biology
The vital a member in identification field.But the finger-image by being shot after infrared light projection by video camera is not only comprising finger vein
Image in itself, in addition to shade caused by the vein tissue of the finger muscles as various thickness, bone and surrounding；Therefore, refer to quiet
Arteries and veins identification technology often faces the problem of accuracy of identification is inadequate；In actual development, refer to hand vein recognition and be often divided into pretreatment, feature carries
Take, characteristic matching, wherein feature extraction is often a most important of which part, extracts the number of feature, quality directly determines
Rate that the match is successful and computational complexity.
Tokyo Univ Japan, South Korea Seoul university in the world, domestic Shandong University, Peking University, Chongqing electronic technology are big
Xue Deng colleges and universities, which all set up right seminar and refer to vein technology, study and has all generated related ends；Propose base
Refer to vein pattern extraction scheme in curve tracing, maximum curvature point, region growing, the detection of wide dimension line, GABOR wave filters etc.,
Major contribution has been made to refer to the development of vein technology, but has referred to vein at present and still has accuracy of identification deficiency, has easily been filled by collection
Influence is put, algorithm redundancy is low, may be only available for the problems such as centralized database.
Existing technical scheme one：To distinguish the feature extraction of bitmap based on personalization, due to traditional LBP feature extractions
Comprising noise, author proposes a kind of personalized differentiation bit map schemes, refers to vein by the difference for contrasting same finger collection
Image, the position of more respective LBP differences, the LBP conducts that final choice goes out non-noise position refer to vein pattern；Prior art side
The feature extraction scheme used in case one is based on personalized LBP, although operand is smaller, accuracy of identification is less than current stream
Row algorithm, redundancy is too low, increases and refuses probability that is true and knowing by mistake, is not suitable for the biological recognition system of high safety.
Existing technical scheme two：To extract finger vein image based on curve tracing, then according to the gray scale of pixel
Value is matched, and image is divided into after fritter stencil matching is used in image to be matched, finds the minimum position of matching error
Put as target location, matching error finally can be obtained to error summation；The feature extraction of two kinds of uses of prior art
Scheme is the extracting method based on curve tracing that Naoto et al. determines by many experiments, based on manual selected characteristic, it is necessary to
Take considerable time and there is no universality completely on different data sets, if changing collecting device, just need to be directed to newly
Refer to vein image re-optimization parameter, extracting method etc..
The content of the invention
It is an object of the invention to overcome the deficiencies in the prior art, and the invention provides a kind of finger vein based on self study
Feature extracting method and system, reduce work difficulty, improve matching precision.
In order to solve the above-mentioned technical problem, the embodiments of the invention provide a kind of finger vein pattern extraction based on self study
Method, the finger vena characteristic extracting method include：
The finger vein image to be identified got is pre-processed, obtains pretreated finger vein image to be identified；
The pretreated finger vein image to be identified is carried out at feature extraction using own coding neural network model
Reason, obtains finger vein pattern to be identified.
Preferably, the described pair of finger vein image to be identified got pre-processes, including：
Gray processing processing is carried out to the finger vein image to be identified, obtains the finger vein image after gray processing processing；
Noise reduction process is carried out to the finger vein image after gray processing processing, obtains the finger vein figure after noise reduction process
Picture；
The size of finger vein image after the noise reduction process is normalized, by the finger after the noise reduction process
The size normalization of vein image is to [60,80].
Preferably, the training process of the own coding neural network model includes：
Own coding neutral net is built, the own coding neutral net includes input layer, hidden layer and output layer；
Suppression processing is carried out to the neuron of the own coding neutral net hidden layer using openness parameter, obtains and suppresses
Own coding neutral net after processing；
Processing is trained using the own coding neutral net after suppression processing described in training sample image set pair, obtained
The own coding neural network model.
Preferably, it is described that the neuron of the own coding neutral net hidden layer is carried out at suppression using openness parameter
Reason, including：
Calculate the neuron average active degree of the own coding neutral net hidden layer；
The openness parameter is made to be equal to the neuron average active degree；
Cause average active value larger with the openness parameter difference in the neuron using extra penalty factor
Less scope is maintained at, obtains the own coding neutral net after suppression processing.
Preferably, the own coding neutral net using after suppression processing described in training sample image set pair is instructed
Practice processing, including：
Connecting quantity and bias node parameter between each layer of the input layer, the hidden layer and the output layer are set
It is set to zero；
Using it is described after training sample image to it is described state suppression processing after own coding neutral net carry out a step training；
Obtain Connecting quantity and the changing value of the bias node parameter between each layer；
According to Connecting quantity and bias node parameter between changing value renewal each layer；
When judging to carry out step training, whether the cost function in training process is less than threshold value, if so, terminate training, if
It is no, a step training step is returned, continues training managing；The threshold value is 3*10-7。
Preferably, described in the use after training sample image to it is described state suppression processing after own coding neutral net enter
The step of row one is trained, including：
Using the residual error of bias node described in the residual sum of Connecting quantity between back-propagation algorithm calculating each layer；
Value is changed according to the residual error of bias node described in Connecting quantity residual sum between each layer for calculating acquisition
Calculate, obtain the changing value of Connecting quantity and the changing value of the bias node between each layer.
Preferably, bias and save described in the residual sum of Connecting quantity between each layer using back-propagation algorithm calculating
The residual error of point, including：
Feedforward conduction is carried out to Connecting quantity and the bias node between each layer to calculate, and obtains the hidden layer
The activation value of activation value and the output layer；
The residual error of each node output unit of the activation value of the output layer is calculated, obtains the output layer
Each node residual error of activation value；
The residual error of each node output unit of the activation value of the hidden layer is calculated, obtains the hidden layer
Each node residual error of activation value；
It is residual according to each node of the activation value of hidden layer described in each node residual sum of the activation value of the output layer
Difference carries out partial derivative calculating, obtains Connecting quantity and the changing value of the bias node parameter between each layer；
The residual error of bias node described in each residual sum of the node residual error including Connecting quantity between each layer.
Preferably, it is described that the pretreated finger vein image to be identified is carried out using own coding neural network model
Feature extraction is handled, including：
The pretreated finger vein image to be identified is inputted in the own coding neural network model；
Pretreated wait to know described by the hidden unit in the hidden layer in the own coding neural network model
Do not refer to vein image diverse location and direction carries out rim detection, extraction refers to vein pattern；
The finger vein pattern extracted is exported by the output layer of the own coding neural network model and represented.
In addition, the embodiment of the present invention additionally provides a kind of finger vein pattern extraction system based on self study, the finger is quiet
Arteries and veins Feature Extraction System includes：
Pretreatment module：For being pre-processed to the finger vein image to be identified got, pretreated treat is obtained
Identification refers to vein image；
Characteristic extracting module：For using own coding neural network model to the pretreated finger vein figure to be identified
As carrying out feature extraction processing, finger vein pattern to be identified is obtained.
Preferably, the characteristic extracting module includes：
Input block：For the pretreated finger vein image to be identified to be inputted into the own coding neutral net mould
In type；
Feature extraction unit：For by the hidden unit in the hidden layer in the own coding neural network model in institute
State pretreated finger vein image diverse location to be identified and direction carries out rim detection, extraction refers to vein pattern；
Feature output represents unit：For the finger vein pattern extracted to be passed through into the own coding neutral net mould
The output layer output of type represents.
In embodiments of the present invention, feature extraction is carried out by the finger vein feature extraction method based on self study, can
In the case where ensuring that finger hand vein recognition precision is enough, to be learnt automatically by machine to the finger with descriptive and high discrimination
Vein pattern, work difficulty is reduced, improve matching precision.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is the required accompanying drawing used in technology description to be briefly described, it is clear that, drawings in the following description are only this
Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, can be with
Other accompanying drawings are obtained according to these accompanying drawings.
Fig. 1 is the method flow schematic diagram of the finger vena characteristic extracting method based on self study in the embodiment of the present invention；
Fig. 2 is the method flow signal of the finger vena characteristic extracting method based on self study in another embodiment of the present invention
Figure；
Fig. 3 is the system architecture composition signal of the finger vein pattern extraction system based on self study in the embodiment of the present invention
Figure；
Fig. 4 is the system architecture composition of the finger vein pattern extraction system based on self study in another embodiment of the present invention
Schematic diagram.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is only part of the embodiment of the present invention, rather than whole embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained all other under the premise of creative work is not made
Embodiment, belong to the scope of protection of the invention.
Embodiment one：
Fig. 1 is the method flow schematic diagram of the finger vena characteristic extracting method based on self study in the embodiment of the present invention,
As shown in figure 1, the finger vena characteristic extracting method includes：
S11：The finger vein image to be identified got is pre-processed, obtains pretreated finger vein figure to be identified
Picture；
S12：Feature is carried out using own coding neural network model to the pretreated finger vein image to be identified to carry
Processing is taken, obtains finger vein pattern to be identified.
S11 is described further：
The finger vein image to be identified got is pre-processed, obtains pretreated finger vein image to be identified.
Specifically, finger vein image to be identified is collected by referring to venous collection device, because similar and different collection
The size for the finger vein image that device collects there may be inconsistent；After finger vein image to be identified is obtained, to be identified
Refer to vein image and carry out gray processing processing, in embodiments of the present invention, the method for gray processing can be component method, maximum value process,
Mean value method or weighted mean method, just do not further provided in the embodiment of the present invention；The each pixel of gray level image is only needed one
Individual byte storage gray value (also known as intensity level, brightness value), tonal range 0-255；Again to be identified after gray processing processing
Refer to vein image and carry out noise reduction, mean filter method may be selected, from appropriate Wiener Filter Method, median filtering method, small echo in noise reduction process
Denoising Algorithm etc., is not construed as limiting in embodiments of the present invention, can select different noise-reduction method as the case may be to gray processing
Finger vein image to be identified carries out noise reduction afterwards；Size is carried out to the finger vein image to be identified after noise reduction using bilinear interpolation to return
One changes, and finger vein image to be identified is uniformly normalized to 60*80 sized images, so when feature extraction, energy
It is quick to carry out feature extraction.
S12 is described further：
The pretreated finger vein image to be identified is carried out at feature extraction using own coding neural network model
Reason, obtains finger vein pattern to be identified.
It is defeated from the input layer in own coding neural network model by the way that pretreated finger vein image to be identified will be carried out
Enter, because the pretreated image size to be identified for referring to vein image is 60*80, that is, pixel be present for 4800, by this
A little pixels are input in 125 hidden units in hidden layer, and network must reconstruct 4800 from 125 dimension hidden layer neurons
The grey scale pixel value x of dimension, some acquisition maximum excitation in 125 hidden units can make it that, it can be seen that different is hidden
Hide unit and carry out rim detection in the different position and direction of image, i.e., study has been arrived extracts feature in specific direction；Will extraction
To feature export and represent by the output layer in own coding neural network model.
In embodiments of the present invention, the training process of own coding neural network model includes：
1st, own coding neutral net is built, the own coding neutral net includes input layer, hidden layer and output layer.
First, according to the actual requirements, an own coding neutral net is built, the neutral net includes input layer, hidden layer
And output layer.
2nd, suppression processing is carried out to the neuron of the own coding neutral net hidden layer using openness parameter, obtains suppression
Own coding neutral net after system processing.
In order to protrude hidden feature, prevent extraneous features to be extracted, enlivened in the present invention by openness suppression neuron
Degree；From sigmoid functions as activation primitive, if it is considered that it is swashed when the output of neuron is close to 1
It is living, think suppressed when output close to 0, then cause all repressed limitation of neuron most of the time to be referred to as openness limit
System.
The process of suppression processing is carried out to the neuron of the own coding neutral net hidden layer such as using openness parameter
Under：
21st, the neuron average active degree of the own coding neutral net hidden layer is calculated；
Especially by the neuron average active degree for calculating own coding neutral net hidden layer, hidden layer neuron j is chosen
Average active degree be：
Wherein m is frequency of training,For given input x(i)Under conditions of, hidden layer neuron j activity.
22nd, the openness parameter is made to be equal to the neuron average active degree；
In order to limit the liveness of hidden layer neuron, order：Wherein ρ is openness parameter, is implemented in the present invention
Example in, openness parameter is arranged to 0.003, can set according to the actual requirements, 0.003 be chosen after many experiments compared with
For preferable parameter.
23rd, average active larger with the openness parameter difference in the neuron is caused using extra penalty factor
Value is maintained at less scope, obtains the own coding neutral net after suppression processing.
In order to realize this limitation, an extra penalty factor will be added in optimization object function so that hide nerve
The average active value larger with difference is maintained at smaller range in member, in the present invention from following form：
Wherein, s2 is the quantity of hidden neuron in hidden layer, and index j represents each neuron.
3rd, processing is trained using the own coding neutral net after suppression processing described in training sample image set pair, obtained
Take the own coding neural network model.
Herein, treat training sample image first to be pre-processed, pretreatment includes gray processing processing, noise reduction process
And normalized, treat that training sample image is classified as gathering by pretreated.Obtain sample image set { (x(1),y(1)),...,(x(i),y(i)), wherein i=1,2 ..., m, the process entirely trained in invention is：All training samples are calculated
A time propagated forward, average activation value is obtained, the activation value of propagated forward in test be out to be temporarily stored in into internal memory and calculated flat
Average.Back-propagating calculating then is carried out to all training samples, and stochastic gradient descent is done to object function, in this hair
It is bright middle to optimize cost function using (L-BFGS) algorithm.
Detailed process is as follows：
31st, Connecting quantity between each layer of the input layer, the hidden layer and the output layer and bias node are joined
Number is arranged to zero.
To all number of plies l (l ∈ 1,2 ..., n), because there was only three layers in neutral net, i.e., in own coding neutral net
Input layer, hidden layer and output layer, so n=3 here in embodiments of the present invention, by each parameterWithInitialization
(normal distribution for the use of standard deviation be 0.01 in the present invention generates random value to random value for very little, close to zero, and order is weighed
Value changes value Δ W(l):=0, Δ b(l):=0, W and b represent Connecting quantity and bias node parameter between each layer of each layer respectively,The Connecting quantity of l layer j units and l+1 layer i units is represented,Represent the bias term of bottom l+1 layer i-th cells.
32nd, using it is described after training sample image to it is described state suppression processing after own coding neutral net carry out a step instruction
Practice, obtain Connecting quantity and the changing value of the bias node parameter between each layer.
Using in above-mentioned sample set, to i=1,2 ..., m sample carry out a step training managing, obtain each layer it
Between Connecting quantity and the changing value of bias node parameter, so as to undated parameter, wherein between each layer Connecting quantity parameter change
It is worth Δ W(l), the changing value Δ b of bias node parameter(l)。
After training sample image to own coding neutral net one step of progress stated after suppression is handled described in the use
The step of training, also includes：
321st, using the residual of bias node described in the residual sum of Connecting quantity between back-propagation algorithm calculating each layer
Difference.
Calculated using back-propagation algorithmWithBack-propagation algorithm：Give
Surely sample (x, y) is inputted, " forward conduction " computing is carried out first, calculates activation value all in network, including hW,b(x)
Output valve；Afterwards to each node i on l layers, calculate its " residual error "The residual error indicates the node to output
The influence of the residual error of value.For final output node, the gap between activation value and actual value caused by network is directly calculated.
For hidden unit, the weighted average based on node residual error calculates input residual error；Detailed process is：
3211st, feedforward conduction is carried out to Connecting quantity and the bias node between each layer to calculate, and is obtained described hidden
Hide the activation value of layer and the activation value of the output layer.
Feedforward conduction is carried out to Connecting quantity between each layer and bias node to calculate, and obtains between each layer Connecting quantity and partially
Put the activation value of node.
3212nd, the residual error of each node output unit of the activation value of the output layer is calculated, obtained described defeated
Go out each node residual error of the activation value of layer.
To each output unit i of each layer (hidden layer and output layer), residual computations, formula are carried out according to following company
It is as follows：
Wherein f (z) represents sigmoid functions, represents vector product, a(nl)The n-th l layer unit activation values are represented,Table
Show l layer i-th cell weighted inputs and δ(nl)Represent the residual error of the n-th l layers.
3213rd, the residual error of each node output unit of the activation value of the hidden layer is calculated, obtained described hidden
Hide each node residual error of the activation value of layer.
To L=nl-1, n-2 ... ..2 each layer, the residual computations method of i-th of node of l layers is as follows：
Wherein, because in own coding neutral net, it is only necessary to the residual error of each node of hidden layer and output layer is calculated,
Therefore hidden layer is first layer, and output layer is the second layer.
3214th, according to each node residual sum of the activation value of the output layer activation value of hidden layer each section
Point residual error carries out partial derivative calculating, obtains Connecting quantity and the changing value of the bias node parameter between each layer.
It is as follows to calculate local derviation formula：
J represents cost function, in the present invention in order to which the openness protrusion of use refers to vein hidden feature, the overall generation of use
Valency function representation is：
Wherein J (W, b) is actual cost function, and β is the weight for controlling openness penalty factor；Item depends on W, b；
Wherein：
Be one using ρ as average and one withFor the relative entropy between two Bernoulli random variables of average.
The residual error of bias node described in each residual sum of the node residual error including Connecting quantity between each layer.
322nd, the residual error of bias node is become according to Connecting quantity residual sum between each layer for calculating acquisition
Change value calculates, and obtains the changing value of Connecting quantity and the changing value of the bias node between each layer.
Calculated according to the residual error of the residual sum bias node of Connecting quantity between each layer of above-mentioned acquisition, calculation formula
It is as follows：
4th, according to Connecting quantity and bias node parameter between changing value renewal each layer.
According to Connecting quantity between each layer of above-mentioned acquisition and the changing value of bias node parameter renewal weight parameter, renewal
Formula is as follows：
When the 5th, judging to carry out step training, whether the cost function in training process is less than threshold value, if so, terminate training,
If it is not, returning to a step training step, continue training managing；The threshold value is 3*10-7。
When judging the training of an above-mentioned step, whether the cost function that uses in the training process is less than threshold value, or whether
Reach maximum iteration set in advance, the threshold size wherein set in the embodiment of the present invention is 3*10-7, in training process
Cost function whether be less than threshold value, if so, terminate training, if it is not, return a step training step, continue training managing.
In embodiments of the present invention, feature extraction is carried out by the finger vein feature extraction method based on self study, can
In the case where ensuring that finger hand vein recognition precision is enough, to be learnt automatically by machine to the finger with descriptive and high discrimination
Vein pattern, work difficulty is reduced, improve matching precision.
Embodiment two：
Fig. 2 is the method flow signal of the finger vena characteristic extracting method based on self study in another embodiment of the present invention
Figure, as shown in Fig. 2 the finger vena characteristic extracting method includes：
S21：Gray processing processing is carried out to the finger vein image to be identified, obtains the finger vein figure after gray processing processing
Picture；
Finger vein image to be identified is collected by referring to venous collection device, because similar and different harvester collection
To finger vein image size there may be it is inconsistent；After finger vein image to be identified is obtained, to finger vein figure to be identified
As carrying out gray processing processing, in embodiments of the present invention, the method for gray processing can be component method, maximum value process, mean value method
Or weighted mean method, just do not further provided in the embodiment of the present invention；A byte is only needed to deposit each pixel of gray level image
Put gray value (also known as intensity level, brightness value), tonal range 0-255.
S22：Noise reduction process is carried out to the finger vein image after gray processing processing, obtains the finger vein after noise reduction process
Image；
Carry out noise reduction to the finger vein image to be identified after gray processing processing, noise reduction process may be selected mean filter method, from
Appropriate Wiener Filter Method, median filtering method, Wavelet-denoising Method etc., are not construed as limiting in embodiments of the present invention, can be according to specific feelings
Condition selectes different noise-reduction method and carries out noise reduction to finger vein image to be identified after gray processing.
S23：The size of finger vein image after the noise reduction process is normalized, after the noise reduction process
Finger vein image size normalization to [60,80].
Size normalization is carried out using bilinear interpolation to the finger vein image to be identified after noise reduction, finger to be identified is quiet
Arteries and veins image uniformly normalizes to 60*80 sized images, so when feature extraction, can quickly carry out feature extraction；Through
Cross after above three step, complete the pretreatment to finger vein image to be identified.
S24：The pretreated finger vein image to be identified is inputted in the own coding neural network model；
It is defeated from the input layer in own coding neural network model by the way that pretreated finger vein image to be identified will be carried out
Enter, because the pretreated image size to be identified for referring to vein image is 60*80, that is, pixel be present for 4800, by this
A little pixels are input in 125 hidden units in hidden layer.
S25：By the hidden unit in the hidden layer in the own coding neural network model described pretreated
Finger vein image diverse location to be identified and direction carry out rim detection, and extraction refers to vein pattern；
Network must reconstruct the grey scale pixel value x of 4800 dimensions from 125 dimension hidden layer neurons, can cause 125 it is hidden
Some hidden in unit obtains maximum excitation, it can be seen that different hidden units enters in the different position and direction of image
Row rim detection, i.e. study have arrived extracts feature in specific direction.
S26：The output layer output table that the finger vein pattern extracted is passed through into the own coding neural network model
Show.
The feature extracted is exported and represented by the output layer in own coding neural network model.
In embodiments of the present invention, feature extraction is carried out by the finger vein feature extraction method based on self study, can
In the case where ensuring that finger hand vein recognition precision is enough, to be learnt automatically by machine to the finger with descriptive and high discrimination
Vein pattern, work difficulty is reduced, improve matching precision.
Embodiment three：
Fig. 3 is the system architecture composition signal of the finger vein pattern extraction system based on self study in the embodiment of the present invention
Figure, as shown in figure 3, the finger vein pattern extraction system includes：
Pretreatment module 11：For being pre-processed to the finger vein image to be identified got, obtain pretreated
Finger vein image to be identified；
Characteristic extracting module 12：For using own coding neural network model to the pretreated finger vein to be identified
Image carries out feature extraction processing, obtains finger vein pattern to be identified.
Preferably, the training process of the own coding neural network model includes：
Own coding neutral net is built, the own coding neutral net includes input layer, hidden layer and output layer；
Suppression processing is carried out to the neuron of the own coding neutral net hidden layer using openness parameter, obtains and suppresses
Own coding neutral net after processing；
Processing is trained using the own coding neutral net after suppression processing described in training sample image set pair, obtained
The own coding neural network model.
Preferably, it is described that the neuron of the own coding neutral net hidden layer is carried out at suppression using openness parameter
Reason, including：
Calculate the neuron average active degree of the own coding neutral net hidden layer；
The openness parameter is made to be equal to the neuron average active degree；
Cause average active value larger with the openness parameter difference in the neuron using extra penalty factor
Less scope is maintained at, obtains the own coding neutral net after suppression processing.
Preferably, the own coding neutral net using after suppression processing described in training sample image set pair is instructed
Practice processing, including：
Connecting quantity and bias node parameter between each layer of the input layer, the hidden layer and the output layer are set
It is set to zero；
Using it is described after training sample image to it is described state suppression processing after own coding neutral net carry out a step training；
Obtain Connecting quantity and the changing value of the bias node parameter between each layer；
According to Connecting quantity and bias node parameter between changing value renewal each layer；
When judging to carry out step training, whether the cost function in training process is less than threshold value, if so, terminate training, if
It is no, a step training step is returned, continues training managing；The threshold value is 3*10-7。
Preferably, described in the use after training sample image to it is described state suppression processing after own coding neutral net enter
The step of row one is trained, including：
Using the residual error of bias node described in the residual sum of Connecting quantity between back-propagation algorithm calculating each layer；
Value is changed according to the residual error of bias node described in Connecting quantity residual sum between each layer for calculating acquisition
Calculate, obtain the changing value of Connecting quantity and the changing value of the bias node between each layer.
Preferably, bias and save described in the residual sum of Connecting quantity between each layer using back-propagation algorithm calculating
The residual error of point, including：
Feedforward conduction is carried out to Connecting quantity and the bias node between each layer to calculate, and obtains the hidden layer
The activation value of activation value and the output layer；
The residual error of each node output unit of the activation value of the output layer is calculated, obtains the output layer
Each node residual error of activation value；
The residual error of each node output unit of the activation value of the hidden layer is calculated, obtains the hidden layer
Each node residual error of activation value；
It is residual according to each node of the activation value of hidden layer described in each node residual sum of the activation value of the output layer
Difference carries out partial derivative calculating, obtains Connecting quantity and the changing value of the bias node parameter between each layer；
The residual error of bias node described in each residual sum of the node residual error including Connecting quantity between each layer.
Specifically, the operation principle of the system related functions module of the embodiment of the present invention can be found in the correlation of embodiment of the method
Description, is repeated no more here.
In embodiments of the present invention, feature extraction is carried out by the finger vein feature extraction method based on self study, can
In the case where ensuring that finger hand vein recognition precision is enough, to be learnt automatically by machine to the finger with descriptive and high discrimination
Vein pattern, work difficulty is reduced, improve matching precision.
Example IV：
Fig. 4 is the system architecture composition of the finger vein pattern extraction system based on self study in another embodiment of the present invention
Schematic diagram, as shown in figure 4, the finger vein pattern extraction system includes：
Gray proces unit：For carrying out gray processing processing to the finger vein image to be identified, gray processing processing is obtained
Finger vein image afterwards；
Noise reduction processing unit：For carrying out noise reduction process to the finger vein image after gray processing processing, noise reduction is obtained
Finger vein image after processing；
Normalized unit：For the size of the finger vein image after the noise reduction process to be normalized,
By the size normalization of the finger vein image after the noise reduction process to [60,80]；
Input block：For the pretreated finger vein image to be identified to be inputted into the own coding neutral net mould
In type；
Feature extraction unit：For by the hidden unit in the hidden layer in the own coding neural network model in institute
State pretreated finger vein image diverse location to be identified and direction carries out rim detection, extraction refers to vein pattern；
Feature output represents unit：For the finger vein pattern extracted to be passed through into the own coding neutral net mould
The output layer output of type represents.
Specifically, the operation principle of the system related functions module of the embodiment of the present invention can be found in the correlation of embodiment of the method
Description, is repeated no more here.
In embodiments of the present invention, feature extraction is carried out by the finger vein feature extraction method based on self study, can
In the case where ensuring that finger hand vein recognition precision is enough, to be learnt automatically by machine to the finger with descriptive and high discrimination
Vein pattern, work difficulty is reduced, improve matching precision.
One of ordinary skill in the art will appreciate that all or part of step in the various methods of above-described embodiment is can
To instruct the hardware of correlation to complete by program, the program can be stored in a computer-readable recording medium, storage
Medium can include：Read-only storage (ROM, Read Only Memory), random access memory (RAM, Random
Access Memory), disk or CD etc..
In addition, a kind of finger vena characteristic extracting method based on self study for being provided above the embodiment of the present invention and being
System is described in detail, and should employ specific case herein and the principle and embodiment of the present invention are set forth, with
The explanation of upper embodiment is only intended to help the method and its core concept for understanding the present invention；Meanwhile for the general of this area
Technical staff, according to the thought of the present invention, there will be changes in specific embodiments and applications, in summary,
This specification content should not be construed as limiting the invention.
Claims (10)
- A kind of 1. finger vena characteristic extracting method based on self study, it is characterised in that the finger vena characteristic extracting method bag Include：The finger vein image to be identified got is pre-processed, obtains pretreated finger vein image to be identified；Feature extraction processing is carried out to the pretreated finger vein image to be identified using own coding neural network model, obtained Take finger vein pattern to be identified.
- 2. the finger vena characteristic extracting method according to claim 1 based on self study, it is characterised in that described pair of acquisition To finger vein image to be identified pre-processed, including：Gray processing processing is carried out to the finger vein image to be identified, obtains the finger vein image after gray processing processing；Noise reduction process is carried out to the finger vein image after gray processing processing, obtains the finger vein image after noise reduction process；The size of finger vein image after the noise reduction process is normalized, by the finger vein after the noise reduction process The size normalization of image is to [60,80].
- 3. the finger vena characteristic extracting method according to claim 1 based on self study, it is characterised in that the own coding The training process of neural network model includes：Own coding neutral net is built, the own coding neutral net includes input layer, hidden layer and output layer；Suppression processing is carried out to the neuron of the own coding neutral net hidden layer using openness parameter, obtains suppression processing Own coding neutral net afterwards；Processing is trained using the own coding neutral net after suppression processing described in training sample image set pair, described in acquisition Own coding neural network model.
- 4. the finger vena characteristic extracting method according to claim 3 based on self study, it is characterised in that described using dilute Dredge property parameter and suppression processing is carried out to the neuron of the own coding neutral net hidden layer, including：Calculate the neuron average active degree of the own coding neutral net hidden layer；The openness parameter is made to be equal to the neuron average active degree；To keep with the larger average active value of the openness parameter difference in the neuron using extra penalty factor In less scope, the own coding neutral net after suppression processing is obtained.
- 5. the finger vena characteristic extracting method according to claim 3 based on self study, it is characterised in that described use is treated Own coding neutral net after suppression processing described in training sample image set pair is trained processing, including：Connecting quantity and bias node parameter between each layer of the input layer, the hidden layer and the output layer are arranged to Zero；Using it is described after training sample image to it is described state suppression processing after own coding neutral net carry out a step training, obtain Connecting quantity and the changing value of the bias node parameter between each layer；According to Connecting quantity and bias node parameter between changing value renewal each layer；When judging to carry out step training, whether the cost function in training process is less than threshold value, if so, terminating training, if it is not, returning A step training step is returned, continues training managing；The threshold value is 3*10-7。
- 6. the finger vena characteristic extracting method according to claim 5 based on self study, it is characterised in that described to use institute State after training sample image to the own coding neutral net one step training of progress stated after suppression is handled, including：Using the residual error of bias node described in the residual sum of Connecting quantity between back-propagation algorithm calculating each layer；Value is changed according to the residual error of bias node described in Connecting quantity residual sum between each layer for calculating acquisition to calculate, Obtain the changing value of Connecting quantity and the changing value of the bias node between each layer.
- 7. the finger vena characteristic extracting method according to claim 6 based on self study, it is characterised in that described using anti- To the residual error of bias node described in the residual sum of Connecting quantity between propagation algorithm calculating each layer, including：Feedforward conduction is carried out to Connecting quantity and the bias node between each layer to calculate, and obtains the activation of the hidden layer The activation value of value and the output layer；The residual error of each node output unit of the activation value of the output layer is calculated, obtains the activation of the output layer Each node residual error of value；The residual error of each node output unit of the activation value of the hidden layer is calculated, obtains the activation of the hidden layer Each node residual error of value；Entered according to each node residual error of the activation value of hidden layer described in each node residual sum of the activation value of the output layer Row partial derivative calculates, and obtains Connecting quantity and the changing value of the bias node parameter between each layer；The residual error of bias node described in each residual sum of the node residual error including Connecting quantity between each layer.
- 8. the finger vena characteristic extracting method according to claim 1 based on self study, it is characterised in that described using certainly Encoding nerve network model carries out feature extraction processing to the pretreated finger vein image to be identified, including：The pretreated finger vein image to be identified is inputted in the own coding neural network model；By the hidden unit in the hidden layer in the own coding neural network model in the pretreated finger to be identified Vein image diverse location and direction carry out rim detection, and extraction refers to vein pattern；The finger vein pattern extracted is exported by the output layer of the own coding neural network model and represented.
- A kind of 9. finger vein pattern extraction system based on self study, it is characterised in that the finger vein pattern extraction system bag Include：Pretreatment module：For being pre-processed to the finger vein image to be identified got, obtain pretreated to be identified Refer to vein image；Characteristic extracting module：For being entered using own coding neural network model to the pretreated finger vein image to be identified Row feature extraction is handled, and obtains finger vein pattern to be identified.
- 10. the finger vein pattern extraction system according to claim 9 based on self study, it is characterised in that the feature Extraction module includes：Input block：For the pretreated finger vein image to be identified to be inputted into the own coding neural network model In；Feature extraction unit：For by the hidden unit in the hidden layer in the own coding neural network model described pre- Finger vein image diverse location to be identified and direction after processing carry out rim detection, and extraction refers to vein pattern；Feature output represents unit：For the finger vein pattern extracted to be passed through into the own coding neural network model Output layer output represents.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710656587.5A CN107563294A (en) | 2017-08-03 | 2017-08-03 | A kind of finger vena characteristic extracting method and system based on self study |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710656587.5A CN107563294A (en) | 2017-08-03 | 2017-08-03 | A kind of finger vena characteristic extracting method and system based on self study |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107563294A true CN107563294A (en) | 2018-01-09 |
Family
ID=60973938
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710656587.5A Pending CN107563294A (en) | 2017-08-03 | 2017-08-03 | A kind of finger vena characteristic extracting method and system based on self study |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107563294A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109919864A (en) * | 2019-02-20 | 2019-06-21 | 重庆邮电大学 | A kind of compression of images cognitive method based on sparse denoising autoencoder network |
| CN110390309A (en) * | 2019-05-22 | 2019-10-29 | 西南科技大学 | A kind of finger vena illegal user's recognition methods based on residual distribution |
| CN111612083A (en) * | 2020-05-26 | 2020-09-01 | 济南博观智能科技有限公司 | Finger vein identification method, device and equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105205449A (en) * | 2015-08-24 | 2015-12-30 | 西安电子科技大学 | Sign language recognition method based on deep learning |
| 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 |
| 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 |
-
2017
- 2017-08-03 CN CN201710656587.5A patent/CN107563294A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105205449A (en) * | 2015-08-24 | 2015-12-30 | 西安电子科技大学 | Sign language recognition method based on deep learning |
| 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 |
| 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 |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109919864A (en) * | 2019-02-20 | 2019-06-21 | 重庆邮电大学 | A kind of compression of images cognitive method based on sparse denoising autoencoder network |
| CN110390309A (en) * | 2019-05-22 | 2019-10-29 | 西南科技大学 | A kind of finger vena illegal user's recognition methods based on residual distribution |
| CN111612083A (en) * | 2020-05-26 | 2020-09-01 | 济南博观智能科技有限公司 | Finger vein identification method, device and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111274916B (en) | Face recognition method and face recognition device | |
| CN112052886A (en) | Human body action attitude intelligent estimation method and device based on convolutional neural network | |
| CN110210555A (en) | Rail fish scale hurt detection method based on deep learning | |
| CN106408562A (en) | Fundus image retinal vessel segmentation method and system based on deep learning | |
| CN107292250A (en) | A kind of gait recognition method based on deep neural network | |
| CN108171318A (en) | One kind is based on the convolutional neural networks integrated approach of simulated annealing-Gaussian function | |
| CN113205538A (en) | Blood vessel image segmentation method and device based on CRDNet | |
| CN108985231B (en) | Palm vein feature extraction method based on multi-scale convolution kernel | |
| CN107392114A (en) | A kind of finger vein identification method and system based on neural network model | |
| CN106909938A (en) | Viewing angle independence Activity recognition method based on deep learning network | |
| CN107563294A (en) | A kind of finger vena characteristic extracting method and system based on self study | |
| CN110210493B (en) | Contour detection method and system based on non-classical receptive field modulation neural network | |
| CN110598552A (en) | Expression recognition method based on improved particle swarm optimization convolutional neural network optimization | |
| Puranik et al. | Intelligent Alzheimer's detector using deep learning | |
| Shan et al. | Residual learning of deep convolutional neural networks for image denoising | |
| CN110111426A (en) | A kind of determination method and apparatus in sound separate pattern house | |
| Sharma et al. | Recognition of plant species based on leaf images using multilayer feed forward neural network | |
| Danukusumo et al. | Indonesia ancient temple classification using convolutional neural network | |
| Al Bataineh et al. | Optimal convolutional neural network architecture design using clonal selection algorithm | |
| Zhang et al. | Application of convolution neural network in Iris recognition technology | |
| CN107944366A (en) | A kind of finger vein identification method and device based on attribute study | |
| CN117218329B (en) | Wellhead valve detection method and device, storage medium and electronic equipment | |
| Wu et al. | RETRACTED ARTICLE: Burn Image Recognition of Medical Images Based on Deep Learning: From CNNs to Advanced Networks | |
| CN107239827A (en) | A kind of spatial information learning method based on artificial neural network | |
| CN108376283A (en) | Pond makeup for neural network is set and pond 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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180109 |
|
| WD01 | Invention patent application deemed withdrawn after publication |