CN109635618A

CN109635618A - Visible images vein developing method based on convolutional neural networks

Info

Publication number: CN109635618A
Application number: CN201810890329.8A
Authority: CN
Inventors: 唐超颖; 押莹; 王彪
Original assignee: Nanjing University of Aeronautics and Astronautics
Current assignee: Nanjing University of Aeronautics and Astronautics
Priority date: 2018-08-07
Filing date: 2018-08-07
Publication date: 2019-04-16
Anticipated expiration: 2038-08-07
Also published as: CN109635618B

Abstract

The invention discloses a visible light image vein imaging method based on a convolutional neural network, which belongs to an information perception and recognition technology. The invention designs an end-to-end full convolutional neural network vein imaging model CY-Net, First, the near-infrared arm images are preprocessed, and then a five-layer convolutional neural network is trained end-to-end with paired visible light arm images and preprocessed NIR images as training samples, and the final network is fully connected. The layer is replaced by a convolutional layer to form a fully convolutional neural network, so that the network can automatically learn the mapping relationship between visible light and near-infrared images, and obtain the target map; Intravenous extraction. Compared with the visible light imaging method based on optics and the three-layer forward neural network, the method of the present invention has better effect, shows better robustness, and verifies the effectiveness of the method of the present invention.

Description

Visible images vein developing method based on convolutional neural networks

Technical field

The invention belongs to a kind of information Perceptions and identification technology, in particular to a kind of visible light based on convolutional neural networks Image vein developing method.

Background technique

In medical domain, intravenous injection is a kind of common treatment means, but the skin of some patients is difficult to find Vein, such as subcutaneous fat is thicker or chaeta is more, vein is relatively thin, the colour of skin is deeper, this makes the success rate of venipuncture not high. Although existing research person develops venous locations detector, traditional hand vein recognition instrument must rely near infrared imaging and set It is standby, it is at high cost, difficult in maintenance.Visible images vein developing method proposed in recent years based on optics and skin physiology can To overcome this limitation (Chaoying Tang, Adams Wai Kin Kong, N.C.Uncovering vein patterns from color skin images for forensic analysis[C]//IEEE International Conference on Computer Vision and Pattern Recognition, 2011:665-672), but the party There is noise, algorithm complexity, to the poor robustness of different skin condition in method.Patent of invention: visible light skin image Vein developing method (application number: 201410734150.5) proposes a kind of visible light imaging based on three layer feedforward neural networks Method, but its structure is too simple, learning ability is weak, therefore imaging results are also limited to.

Deep learning is a kind of method of automatic learning characteristic, wherein convolutional neural networks (Convolutional Neural Networks, CNN) pixel can be retained when handling image spatial relation, its structure is to figure There is height invariance therefore to be very suitable to do the feature of two-dimensional image data for the deformation such as rotation, translation, the scaling of picture It extracts.

The process that vein is extracted from visible light arm image includes that object detection and lines extract in terms of two.Object inspection It surveys, i.e., detects arm from input picture；Lines extract, i.e., the visible light and near-infrared image learnt using network model Mapping relations realized from the arm detected vein image, finally carry out lines extraction again.Traditional object detecting method It is low-level feature using the feature of design, it is insufficient to the ability to express of target.Deep learning method can be from a large amount of sample High-level characteristic is arrived in automatic study, and compared to the feature of engineer, the feature learnt is more abundant, ability to express is also stronger.

With the continuous development of deep learning, researcher's discovery carries out target detection using CNN, and accuracy rate can be mentioned substantially It rises, compared to Conventional visual feature, convolutional neural networks visual signature is more suitable for image detection task.In such as venous blood Pipe, lane, course line lines extract aspect, traditional algorithm is strong to the identification of feature.And the expression way of feature excessively according to Rely artificial selection, researcher not can guarantee the accurate selection and expression to feature.The deep layer group of deep learning main analog human brain Structure is knitted, low-level feature is combined.More abstract effective high-rise expression is formed, therefore mark sheet can be carried out automatically It reaches.

Superfine finger hand vein recognition algorithm (the finger hand vein recognition research based on deep learning proposed based on deep learning of Wu [J] computer technology and development, 2018 (2): 200-204), it proposes to improve on the basis of Alex Net deep neural network Scheme, carrying out classification to finger vein image has better result；Qin Zhi light etc. proposes a kind of based on multi-modal CNN model ((cerebrovascular Study on Extraction Method [J] Journal of UEST of China based on multi-modal convolutional neural networks, 2016,45 (4): 573-581)), brain CT angiographic image is split, and has been experimentally confirmed this model and has been extracted in the cerebrovascular On validity；The it is proposeds such as Wang Yiding obtained using wavelet decomposition and the method for down-sampling it is multiple dimensioned under hand back vein figure Picture, then the extraction of characteristics of image is carried out with centrosymmetric local binary patterns, it is successively trained using limitation Boltzmann machine And discrimination (the hand based on deep learning and multiple dimensioned coded combination of vein is improved by the method for multiple dimensioned coded combination Dorsal vein identifies [J] North China University of Tech journal, 2015,27 (3): 6-13)；Zhang Guiying etc. uses two kinds of nets based on CNN Network structure automatically learns and extracts the feature of optic disk in retinal images, realizes the automatic detection of optic disk, substantially increases view Disk detection accuracy rate (optic disk of Zhang Guiying, the Zhang Xianjie based on deep learning detects the Guizhou [J] college of education journal automatically, 2017,33(3):27-32)。

Convolutional neural networks can extract high-level characteristic, and feature extraction, feature selecting and tagsort are merged same Training, non local on the whole to optimize in model, is enhanced the separability of feature.CNN is mentioned applied to vein Other methods can effectively be overcome the shortcomings of by taking.But there are no the skills that CNN is used for the imaging of visible light vein in the prior art Art.

Summary of the invention

The present invention is directed to traditional vein and images the dependence near infrared gear and obtained in the prior art using optical means To image there are obvious noise, imaging results are limited, and the defect of poor robustness proposes a kind of based on the visible of deep learning CNN is used for the technology of visible light vein imaging, solved by light image vein developing method by devising a kind of CNN model Problems of the prior art.

The present invention is implemented as follows:

The invention discloses a kind of visible images vein developing method based on convolutional neural networks, specific steps are such as Under:

Step 1: near-infrared arm vein image preprocessing；

Step 2: designing end as training sample based on the pretreated image of step 1 and visible light arm image Full convolutional neural networks vein to end images MODEL C Y-Net；The network is using visible light arm image as input, using five Layer convolutional layer is filtered operation to image；

Step 3: to the parameter optimization strategy in CY-Net network model: using gradient optimization algorithm and cost Function optimization algorithm；

Step 4: finally carrying out vein extraction to the vein imaging figure of network output.

Further, the adaptive histogram equalization method near-infrared arm of contrast-limited is utilized in the step one Image is pre-processed.

Further, the step two specifically:

2.1, the convolutional neural networks that the network overall structure of design is one five layers are followed by one behind each convolutional layer Layer PReLU activation primitive layer, common activation primitive Sigmoid and Tanh is saturation nonlinearity function.In recent years, most nerves Network all uses unsaturated ReLU (Rectified Linear Unit, non-liner revision unit) or its deformation PReLU The function of (Parametric Rectified Linear Unit) as active coating；Its mathematical expression difference is as follows:

F (x)=max (0, x) (1)

Wherein, x is function input, and f (x) is function output, and a is slope；

One layer of PReLU activation primitive layer is followed by behind each convolutional layer of present networks, increase network non-linear factor and The learning ability of model, it, which increases a small amount of calculation amount than ReLU activation primitive, realizes higher accuracy rate, believes negative axis Breath will not all be lost, and avoid " dying ReLU ".

2.2, Dropout layers are arranged after convolutional layer four, the characteristic pattern finally obtained with 3 × 3 convolution kernel to convolutional layer 5 It carries out convolution and obtains network output image, calculate the cost function of the near-infrared image of output image and input, passed using reversed The gradient descent method optimization that iterates is broadcast to update network parameter and minimize cost function, make e-learning to visible images with Mapping relations between near-infrared image；

Dropout layers be a kind of generalization ability for avoiding occurring in network training over-fitting and improving network plan Slightly, network complexity is reduced by giving up the output of hidden layer neuron at random.The neuron being rejected both be not involved in front of to Communication process is also not involved in back-propagation process, can reduce the coupling between neuron in this way, make network have more generalization ability, It is more robust.When training data is less, it is easy to happen over-fitting, Dropout strategy can be used.

2.3, using pretreated image as supervision message, one five layers of full convolutional Neural net is trained end-to-endly Network, direct prediction label figure；The target of end-to-end study is directly learnt by depth network to (defeated from the primitive form of data Enter) to the mapping relations between the label (output) of data, reduce artificial pretreatment and post-processing and complicated based on artificial The rule of design or the feature of selection increase the whole compatible degree of network.

It is specially input with visible images, corresponding near-infrared image is that target exports, trained one five layers complete Convolutional neural networks；Network output is subjected to backpropagation as error at a distance from target output, is optimized and revised in network The parameters such as weight, bias term, reduce error constantly, until network output is exported close to target.

Further, first and third layer of convolution kernel size is designed as 5, remaining is 3.Convolution operation usually has and makes Useful information enhancing, the effect for reducing noise are practice have shown that lesser convolution kernel preferably extracts the feature in image.

Further, cost function optimization algorithm is specially the model for using L2 regularization to carry out constrained parameters in the step three Number, i.e., with the difference of predicted value and actual value square half to being averaged after the summation of m sample come measure algorithm operating condition:

In formula, W is network weight, and b is network bias term, and m is number of samples,It is network to the mesh of i-th sample Mark output, y⁽ⁱ⁾It is network to the reality output of i-th of sample,

Therefore the final goal of Solution To The Network is that the minimum value of cost function, solution procedure are asked for weight W and bias term b Gaussian initialization is carried out to each parameter；

L2 regularization is substantially linear attenuation to be carried out to the weight of network, therefore L2 regularization is also referred to as weight decaying, It is the cost function formula (3) in network below along with a regularization term, it may be assumed that

In formula, the subsequent Section 2 of plus sige is regularization term；

L2 regularization term be to all parameter W square summation after be averaged multiplied by λ/2, wherein λ is regularization coefficient, For weighing specific gravity shared by original cost function J (W, b) and regular terms plus derivation again after regular terms, the more new formula of W Become:

In formula, Section 2 is the partial derivative for solving parameter, and α indicates learning rate；

So the coefficient of W becomes the value smaller than 1 from 1 in newerIts effect is to reduce W to update weight Influence, i.e., weight decay.

Further, gradient optimization algorithm carries out parameter update to weight and bias term in the step three, updates Mode is as follows:

In formula,Indicate the weight parameter connecting between j-th of node of l layer and l+1 i-th of node of layer, Indicate the bias term of l+1 i-th of node of layer, α indicates learning rate, can be seen that by two parameter more new formulas and solves parameter Partial derivative is the key that gradient descent method；Cost function is to weightAnd bias termLocal derviation is asked to obtain:

Residual error is to calculate the output node of every layer network the difference of output valve and true value, it is assumed that the output layer of neural network It is n-th_lLayer, then the residual error of output layer is usedIt indicates, l=n_l-1,n_lThe residual error of i-th of node of -2, K, 2 layer is usedTable Show, calculation formula is as follows:

In formula,It is got by the weighted average calculating of l+1 node layer residual error, after having two above residual error, then Cost function J (W, b) can be further calculated to the local derviation of weight and bias term, it may be assumed that

Detailed description of the invention

Fig. 1 is near-infrared arm vein image preprocessing result figure of the present invention；

Fig. 2 is the pretreatment schematic diagram of arm near-infrared image of the present invention；

Fig. 3 is the schematic diagram of CY-Net network structure of the present invention；

Fig. 4 is the image of arm vein post processing of image process of the present invention；

Fig. 5 is to show in the embodiment of the present invention using visible images, based on what optical means and method of the invention obtained As comparative result figure；

Fig. 6 is the local binarization image after vein of the present invention extracts；

Fig. 7 is network in the embodiment of the present invention to the effect picture of data difference fitting degree；

Fig. 8 is the preferable three groups of images of subjective assessment in the embodiment of the present invention and its result figure after binaryzation；

Fig. 9 is CY-Net model robustness verification result figure under shooting conditions different in the embodiment of the present invention.

Specific embodiment

It is clear to keep the purpose of the present invention, technical solution and effect clearer, referring to the drawings and illustrative example pair The present invention is further described.It should be understood that specific implementation described herein is not used to limit only to explain the present invention The fixed present invention.

1. near-infrared image pre-processes

Due to the specific location of near-infrared arm image medium sized vein and unintelligible, utilize the adaptive of contrast-limited Histogram equalization method (Contrast Limited Adaptive Histogram Equalization, CLAHE) is answered to mention The contrast of near-infrared image is risen, as shown in Figure 1.1 (a) is arm near-infrared image, and 1 (b) is the adaptive of contrast-limited The result of histogram equalization method.

The data set of network is made of the synchronous arm image of visible light-near-infrared of 24 people, since sample total is less, Therefore final to obtain 96 pairs of visible lights-near-infrared arm image using horizontal, vertical mirror method EDS extended data set, at random Wherein 80% training set of the data as network is chosen, remaining 20% data is as test set.

2. full convolutional neural networks vein images MODEL C Y-Net end to end

The CY-Net network structure that the present invention designs is as shown in Figure 3: the network is adopted using visible light arm image as input Operation, convolution kernel size and feature map number are filtered to image with five layers of convolutional layer as shown in figure 3, after each convolutional layer It is followed by one layer of PReLU activation primitive layer, Dropout layers of setting inactivates neuron at random excessively quasi- to avoid network after convolutional layer 4 It closes, convolution finally is carried out to the characteristic pattern that convolutional layer 5 obtains with 3 × 3 convolution kernel and obtains network output image, calculates output figure As the cost function with the near-infrared image of input, iterating to optimize using back-propagation gradient descent method updates network parameter Cost function is minimized, makes the e-learning to the mapping relations between visible images and near-infrared image.

First and third layer of convolution kernel size is designed as 5, remaining is 3.Convolutional layer is the core layer in convolutional neural networks, Convolutional filtering operation is carried out to each channel of input picture, extracts different features.The each convolutional layer of network of the invention It is followed by one layer of PReLU activation primitive layer below, increases the non-linear factor of network and the learning ability of model, it swashs than ReLU Function living, which increases a small amount of calculation amount, realizes higher accuracy rate, lose negative axis information will not all, avoids " dying ReLU".It deforms the function of PReLU (Parametric Rectified Linear Unit) as active coating, Mathematical expression difference is as follows:

F (x)=max (0, x) (1)

In formula, x is function input, and f (x) is function output, and a is slope.

Dropout layers: the coupling between neuron is reduced by giving up hidden layer neuron at random, to reduce network Complexity makes network have more generalization ability, more robust.In order to avoid generation over-fitting, the generalization ability of network is improved, Dropout strategy is used when network design.

Finally the full articulamentum with output connection realizes classification task for extracting high-level characteristic to traditional CNN network.This Invention is based on the theory of FCN (Fully Convolutional Network), by full connection last in traditional CNN network Layer has changed convolutional layer into, constitutes full convolutional network, obtains the mapping graph of target finally to realize the imaging of vein image.

The CNN network C Y-Net that the present invention designs is with CLAHE treated near-infrared arm image (Ground truth) As supervision message, full convolutional neural networks of one five layers of training end-to-endly, direct prediction label figure (Label map), As shown in Fig. 2, being input with visible images, corresponding near-infrared image is target output, the full convolution of one five layers of training Neural network.By network output with as error progress backpropagation, optimized and revised at a distance from target output weight in network, The parameters such as bias term, reduce error constantly, until network output is exported close to target.

3. Network Optimization Strategy

(1) cost function optimizes

The present invention is using L2 regularization come the norm of constrained parameters.Vein visualization role substantially belongs to recurrence task, because The cost function of this network is defined as in recurrence task common Euclidean distance loss function, that is, uses the difference of predicted value and actual value Square half to being averaged after the summation of a sample and carry out measure algorithm operating condition:

Therefore the final goal of Solution To The Network is that the minimum value of cost function is sought for weight W and bias term b.Solution procedure Gaussian initialization is carried out to each parameter.

L2 regularization is substantially linear attenuation to be carried out to the weight of network, therefore L2 regularization is also referred to as weight decaying (weight decay).It is the cost function formula (3) in network below along with a regularization term, it may be assumed that

L2 regularization term be to all parameter W square summation after be averaged multiplied by λ/2, wherein λ is regularization coefficient, For weighing specific gravity shared by original cost function J (W, b) and regular terms.In addition derivation again after regular terms, the more new formula of W Become:

So the coefficient of W becomes the value smaller than 1 from 1 in newerIts effect is to reduce W to update weight Influence, i.e., weight decay.In a sense, smaller weight indicates that the complexity of network is lower, therefore, pass through constraint The norm of parameter, regularization can mitigate over-fitting to a certain extent.

(2) gradient descent method

The present invention mainly uses Adam (Adaptive Moment Estimation) optimization algorithm.It is stochastic gradient The extension form of descent algorithm (Stochastic Gradient Descent, SGD), algorithm fusion AdaGrad and The advantages of RMSProp algorithm, is based on first moment mean value computation adaptability parameter learning rate, and makes full use of the second moment of gradient equal Value.According to practical experience, the exponential damping parameter beta1 of single order moments estimation takes 0.9, the exponential damping parameter of second order moments estimation Beta2 takes 0.999, epsilon parameter to take 10E-8.

Gradient optimization algorithm carries out parameter update to weight and bias term in step 3, and update mode is as follows:

Ubuntu16.04, the Caffe deep learning frame of 64 bit manipulation systems, with NVIDIA are based in the present embodiment TITAN X (Pascal), the GPU of 8G memory are experimental situation.Vein image is shot using JAI-AD080CL camera, the camera The spectrum for obtaining visible light and near-infrared can be synchronized.By convolutional neural networks MODEL C Y-Net complete end to end proposed by the present invention What is proposed in the prior art is compared based on optical developing method, to verify the validity of network.Two methods are obtained To output image and former near-infrared arm image post-processed.

Firstly, carrying out Gabor filtering to obtained vein image, Gabor filter can automatically extract the quiet of output image Then based on this arteries and veins information increases venous information using the energy diagram exported after arm shape exposure mask and Gabor filtering By force, binaryzation is finally carried out to enhanced image using the method for adaptive threshold.To the process of arm NIR post processing of image As shown in Figure 4.4 (a) be NIR image；4 (b) be Gabor filtering and enhanced result；4 (c) be the result after binaryzation.

It subjective assessment based on the method for the present invention and is analyzed as follows:

As shown in figure 5, Fig. 5 is that the result that the visible images of arm, optical means obtain and CY-Net are obtained respectively As a result.The imaging of visible light vein image can be achieved in optical means and CY-Net as seen from Figure 5, wherein optical means needle Each pixel is handled, there are obvious noises, and the method based on deep learning is end-to-end study, and noise is than optics side Method wants small.Fig. 6 is the local binarization figure of the 9th group of result, further demonstrates the vein extracting method based on deep learning and makes an uproar Sound is smaller.From the 5th, 7, the 8 group of result of Fig. 5 also it will be evident that the result obtained based on optical method is in dark Region all pixels gray value is all very low, can not observe venous locations, and the method based on deep learning overcomes this and asks Topic, embodies preferable robustness.

It objectively evaluating based on the method for the present invention and is analyzed as follows:

In order to further verify the validity of invention, it is objective to be carried out with four common counters in identification field to experimental result Evaluation: accuracy rate (Accuracy), precision (Precision), also known as precision ratio, recall ratio (Recall) and FM estimate.It is fixed Justice is as follows respectively:

By definition it is found that accuracy rate and precision ratio measured is accuracy that algorithm extracts vein, recall ratio is then paid close attention to Be extract result comprehensive, FM, which estimates, carries out overall merit to algorithm performance in conjunction with precision ratio and recall ratio, to three kinds of sides The bianry image that method post-processes, aforementioned four index value is bigger to illustrate that vein extracts result closer to true vein, effect Better.Vein extraction optically is carried out with CY-Net respectively on 20 groups of visible light arm images, carries out post-processing it Afterwards, four evaluation indexes for calculating every group of result, using near-infrared image corresponding with visible images as Calculation Estimation index True value (Ground truth).It is the accuracy rate and precision ratio evaluation index value of two methods in table 1, is two kinds of sides in table 2 The recall ratio and FM measuring and evaluation index value of method.

Table 1 is Accuracy the and Precision evaluation index value of two methods

Table 2 is Recall the and FM evaluation index value of two methods

Data comparison in table it can be found that in 20 groups of test images, CY-Net have 70% Accuracy value and Precision value is higher than optical means, and the average value on Accuracy, Precision and FM estimate is all than being based on optics Method it is high, illustrate that CY-Net can reach effect more better than optical means based on Pixel-level.Network proposed by the present invention is tested Having demonstrate,proved convolutional neural networks has the ability to the substantive characteristics of image that learns from low volume data.

Fitting degree based on the method for the present invention is analyzed as follows:

Poor fitting refers to that model does not capture the feature of data well, can not good fitting data；Over-fitting refers to Model is fitted too thorough data, so that the feature of data noise is also learnt to arrive, will lead to can not in test Data are identified well, and the generalization ability of model is poor.As shown in fig. 7, being from left to right trained poor fitting (a) respectively, preferably quasi- The case where closing (b) and over-fitting (c).

Wherein Fig. 7 (b) is the visitor by the preferable three groups of results of subjective assessment in measurement same model, different the number of iterations See what index obtained.Fig. 8 is the preferable three groups of images of subjective assessment, calculates four kinds of evaluation indexes, knot to Fig. 8 (a), (b), (c) Fruit is shown in Table 3, each evaluation index of Comprehensive Correlation.It is considered that (c) fitting degree in Fig. 8 is preferable, as used in inventive network Model.

Table 3 is four kinds of evaluation index values of tri- groups of results of Fig. 8

Training starts, and model is often poor fitting, also just there is the space of optimization just because of this, can be by constantly The hyper parameters such as regularized learning algorithm rate and increase the number of iterations to make the feature of model more fully learning data, makes the expression of model Ability is stronger.The reason of over-fitting occurs also has very much, usually will appear over-fitting in the case where data set is less, can To expand data set, therefore the data set of script 4 times are expanded；Meanwhile the number of iterations is not The more the better, sometimes Big the number of iterations model learning has arrived the details of overabundance of data, effect be not so good as instead the number of iterations it is relatively fewer when model, Therefore, suitable the number of iterations is selected to can also be improved the expression quality of model；It, can also in order to which over-fitting is alleviated or avoided The cost function and Dropout strategy that network is defined using regularization method, by changing network structure, reduction neuron Coupling improve the generalization ability of model.

Robustness verifying based on the method for the present invention is as follows:

Since visible light-near-infrared arm image that network training is concentrated into pair is shot by camera of the same race, illumination condition, The variation of the factors such as the receptance function of camera is little.Therefore the experiment of this part is mainly verified in different illumination conditions, is clapped It takes the photograph under environment, and to the robustness of non-arm position network model.As shown in figure 9, CY-Net mould under Fig. 9 difference shooting condition The verifying of type robustness.Shown in Fig. 10 is different physical feeling CY-Net model robustness verifyings；The left side of every a pair of image is Visible images, the right are CY-Net vein imaging figure.

Have been verified that model proposed by the present invention has preferable Shandong to illumination variation in subjective assessment and analysis part Stick.From Fig. 9 and Figure 10 can be seen that the model that the method based on deep learning trains learn automatically to visible light with it is close red Complicated mapping relations between outer image, it is effective not only for the arm image of training sample, change in external condition or All there is preferable robustness and generalization ability when image content change itself.

The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For member, without departing from the principle of the present invention, several improvement can also be made, these improvement also should be regarded as of the invention Protection scope.

Claims

1. a visible light image vein imaging method based on convolutional neural network, is characterized in that, step is as follows:

Step 1. Preprocessing of near-infrared arm vein images;

Step 2: Design an end-to-end fully convolutional neural network vein imaging model CY-Net based on the preprocessed images and visible light arm images in step 1 as training samples;

Step 3. Parameter optimization strategy in the CY-Net network model: use gradient descent optimization algorithm and cost function optimization algorithm;

Step 4: Finally, perform vein extraction on the vein imaging image output by the network.

2. a kind of visible light image vein imaging method based on convolutional neural network according to claim 1, is characterized in that, utilizes contrast-limited adaptive histogram equalization method near infrared arm image in described step 1 preprocessing.

3. a kind of visible light image vein imaging method based on convolutional neural network according to claim 1, is characterized in that, described step 2 is specifically:

2.1. The overall structure of the designed network is a five-layer convolutional neural network, and each convolutional layer is followed by a layer of PReLU activation function; its mathematical expressions are as follows:

f(x)=max(0,x) (1)

Among them, x is the function input, f(x) is the function output, and a is the slope;

2.2, set the Dropout layer after the convolutional layer 4, and finally use the 3×3 convolution kernel to convolve the feature map obtained by the convolutional layer 5 to obtain the network output image, and calculate the cost function of the output image and the input near-infrared image, Use the back-propagation gradient descent method to iteratively optimize and update the network parameters to minimize the cost function, so that the network can learn the mapping relationship between the visible light image and the near-infrared image;

2.3, take the visible light image as the input and the corresponding near-infrared image as the target output, train a five-layer full convolutional neural network; use the distance between the network output and the target output as the error for back-propagation, and optimize and adjust the parameters in the network , so that the error continues to decrease until the network output is close to the target output.

4. a kind of visible light image vein imaging method based on convolutional neural network according to claim 3, is characterized in that, described first and third layer convolution kernel size is designed to be 5, and the rest are all 3.

5. a kind of visible light image vein imaging method based on convolutional neural network according to claim 1, is characterized in that, in described step 3, cost function optimization algorithm is specifically adopting L2 regularization to constrain the norm of parameter , that is, use the half of the square of the difference between the predicted value and the actual value to sum the m samples and take the average to measure the operation of the algorithm:

In the formula, W is the network weight, b is the network bias term, m is the number of samples, is the target output of the network for the ith sample, y ⁽ⁱ⁾ is the actual output of the network for the ith sample,

Therefore, the ultimate goal of the network solution is to find the minimum value of the cost function for the weight W and the bias term b, and the solution process performs Gaussian initialization on each parameter;

L2 regularization is essentially a linear attenuation of the weight of the network, so L2 regularization is also called weight attenuation, which is to add a regularization term to the cost function of the network (3), that is:

In the formula, the second term after the plus sign is the regularization term;

The L2 regularization term is the sum of the squares of all parameters W and then averaged and then multiplied by λ/2, where λ is the regularization term coefficient, which is used to weigh the original cost function J(W, b) and the proportion of the regularity term , add the regular term and then take the derivative, the update formula of W becomes:

In the formula, the second term is the partial derivative of the solution parameter, and α represents the learning rate;

Therefore, the coefficient of W in the update formula changes from 1 to a value smaller than 1 The effect is to reduce the effect of W on weight updates, ie weight decay.

6. A kind of visible light image vein imaging method based on convolutional neural network according to claim 1, it is characterized in that, in described step 3, gradient descent optimization algorithm carries out parameter update to weight and bias term, update method as follows:

In the formula, represents the weight parameter of the connection between the jth node in the lth layer and the ith node in the l+1th layer, Represents the bias term of the i-th node in the l+1th layer, and α represents the learning rate. From the two parameter update formulas, it can be seen that the partial derivative of the solution parameter is the key to the gradient descent method; and bias term Find the partial derivative:

The residual is the difference between the output value and the real value calculated for the output node of each layer of the network. Assuming that the output layer of the neural network is the n _lth layer, the residual of the output layer is Represents that the residual of the i-th node of the l=n _l -1, n _l -2, K, 2 layers is used said, its calculation formula is as follows:

In the formula, It is calculated by the weighted average of the residuals of the nodes in the l+1 layer. After the above two residuals are obtained, the partial derivatives of the cost function J(W,b) to the weights and bias terms can be further calculated, namely: