Authentication method, device, electronic device and readable storage medium

Abstract

The present application provides an identity authentication method, device, electronic device and readable storage medium, which relate to the field of image recognition technology. The method uses an image recognition model to identify the similarity between the current biometric image and the historical biometric image of the object to be verified. The image recognition model is a twin neural network model, in which two network models are residual networks including Gabor filters. Since the Gabor filter can extract robust features for direction and scale changes in the convolution layer of the residual network, it can better extract image features, and then use the twin neural network model to more accurately identify the similarity of the two images, achieve better recognition effect, and more accurately verify the identity of the object to be verified.

Description

Technical Field

[0001] The present application relates to the field of image recognition technology, and in particular to an identity authentication method, device, electronic device and readable storage medium. Background Art

[0002] With the development of science and technology, recognition technology has been widely used. In the existing method, the user's image information is generally collected for recognition, such as fingerprints, faces, etc. At present, convolutional neural networks are usually used to extract features from enhanced images, and then classified and recognized through classifiers to obtain corresponding recognition results.

[0003] Although the above recognition technology can realize the recognition of the user's identity, since in the above recognition process, only enhanced processing is performed and a classifier is used for classification and recognition, the accuracy of the recognition result is relatively low. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide an identity authentication method, device, electronic device and readable storage medium to improve the problem of low accuracy of the recognition method in the prior art.

[0005] In a first aspect, an embodiment of the present application provides an identity authentication method, the method comprising:

[0006] Acquire a current biometric feature image of the object to be authenticated and acquire a pre-stored historical biometric feature image of the object to be authenticated;

[0007] Using an image recognition model to identify the similarity between the current biometric image and the historical biometric image, and obtaining a recognition result, wherein the image recognition model is a twin neural network model, two network models in the twin neural network model are residual networks including Gabor filters, and the two network models are respectively used to extract image features of the current biometric image and the historical biometric image;

[0008] The identity of the object to be authenticated is authenticated according to the recognition result.

[0009] In the above implementation process, an image recognition model is used to identify the similarity between the current biometric image and the historical biometric image of the object to be verified. The image recognition model is a twin neural network model, and the two network models are residual networks including Gabor filters. Since the Gabor filter can extract robust features against direction and scale changes in the convolutional layer of the residual network, it can better extract image features, and then the twin neural network model can more accurately identify the similarity of the two images, thereby achieving better recognition effect, so as to more accurately verify the identity of the object to be verified.

[0010] Optionally, the twin neural network model further includes a concatenation layer, a fully connected layer, and a classification layer, and the using the image recognition model to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result includes:

[0011] Extracting image features of the current biometric image through one of the two network models to obtain a first feature vector, and extracting image features of the historical biometric image through the other network model to obtain a second feature vector;

[0012] Concatenate the first feature vector and the second feature vector through the concatenation layer to obtain a concatenated vector;

[0013] The series vector is passed through the fully connected layer and the classification layer in sequence to obtain a recognition result.

[0014] In the above implementation process, two network models are used to extract image features, and the feature vectors are concatenated for classification and recognition. In this way, the features including scale and direction extracted by the network model can be effectively utilized, thereby achieving better recognition effects.

[0015] Optionally, the concatenating the first feature vector and the second feature vector through the concatenation layer to obtain a concatenated vector includes:

[0016] calculating the square of the difference between the first eigenvector and the second eigenvector through the concatenated layer, and calculating the Hadamard product between the first eigenvector and the second eigenvector through the concatenated layer;

[0017] The first eigenvector, the second eigenvector, the square of the difference, and the Hadamard product are connected in series through the connection layer to obtain a series connection vector.

[0018] In the above implementation process, by connecting two feature vectors, the square of the difference and the Hadamard product in series, more features can be obtained to participate in the recognition, so that the recognition effect is better.

[0019] Optionally, the recognition result is a probability that both the current biometric image and the historical biometric image belong to the object to be verified, and verifying the identity of the object to be verified according to the recognition result includes:

[0020] If the probability is greater than a set threshold, it is determined that the identity authentication of the object to be authenticated is passed;

[0021] If the probability is less than or equal to the set threshold, it is determined that the identity authentication of the object to be authenticated has failed.

[0022] In the above implementation process, the identity of the object to be verified can be directly and quickly verified through probability.

[0023] Optionally, the model parameters between the two network models are shared, which can make the model training faster and facilitate the implementation of the model.

[0024] Optionally, the biometric image is a fingerprint image, a palm print image, an iris image or a finger vein image, so that the solution can be applied to more recognition scenarios.

[0025] Optionally, when the biometric image is a finger vein image, the step of obtaining the current biometric image of the object to be verified includes:

[0026] The current finger vein images of at least two fingers of the object to be verified are obtained, wherein the historical biometric image includes the historical finger vein images of the at least two fingers. In this way, more images can be collected to participate in the recognition, so as to improve the accuracy of the recognition.

[0027] In a second aspect, an embodiment of the present application provides an identity authentication device, the device comprising:

[0028] An image acquisition module, used to acquire a current biometric image of the object to be authenticated and to acquire a pre-stored historical biometric image of the object to be authenticated;

[0029] A recognition module, used to use an image recognition model to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result, wherein the image recognition model is a twin neural network model, two network models in the twin neural network model are residual networks including Gabor filters, and the two network models are respectively used to extract image features of the current biometric image and the historical biometric image;

[0030] A verification module is used to verify the identity of the object to be verified according to the recognition result.

[0031] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in the method provided in the first aspect are performed.

[0032] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps in the method provided in the first aspect are performed.

[0033] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the embodiments of the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 A flowchart of an identity authentication method provided in an embodiment of the present application;

[0036] Figure 2 (a) is a schematic diagram of the structure of a common residual network provided in an embodiment of the present application;

[0037] Figure 2 (b) is a schematic diagram of the structure of a residual network including a Gabor filter provided in an embodiment of the present application;

[0038] Figure 3 A schematic diagram of the structure of an image recognition model provided in an embodiment of the present application;

[0039] Figure 4 (a) is a schematic diagram of an image in a FV-USM dataset provided in an embodiment of the present application;

[0040] Figure 4 (b) is a schematic diagram of an image in the MMCBNU_6000 data set provided in an embodiment of the present application;

[0041] Figure 5 (a) is a schematic diagram showing a comparison of training times in different directions and scales in a FV-USM dataset provided by an embodiment of the present application;

[0042] Figure 5 (b) is a schematic diagram for comparing training times of different directions and scales in the MMCBNU_6000 dataset provided in an embodiment of the present application;

[0043] Figure 6 (a) is a schematic diagram of the relationship between the size of each model and ACC and EER under the FV-USM dataset provided by an embodiment of the present application;

[0044] Figure 6(b) is a schematic diagram of the relationship between the size of each model and ACC and EER under the MMCBNU_6000 data set provided in an embodiment of the present application;

[0045] Figure 7 (a) is a schematic diagram of the relationship between the size of each model and DET under a FV-USM data set provided in an embodiment of the present application;

[0046] Figure 7 (b) is a schematic diagram of the relationship between the size of each model and DET under a MMCBNU_6000 data set provided in an embodiment of the present application;

[0047] Figure 8 A schematic diagram of the structure of an identity verification device provided in an embodiment of the present application;

[0048] Fig. 9 A schematic diagram of the structure of an electronic device for executing an identity authentication method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application.

[0050] It should be noted that the terms "system" and "network" in the embodiments of the present invention can be used interchangeably. "Multiple" means two or more. In view of this, in the embodiments of the present invention, "multiple" can also be understood as "at least two". "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / ", unless otherwise specified, generally indicates that the previously associated objects are in an "or" relationship.

[0051] An embodiment of the present application provides an identity authentication method, which uses an image recognition model to identify the similarity between a current biometric image and a historical biometric image of an object to be verified. The image recognition model is a twin neural network model, and the two network models are residual networks including Gabor filters. Since the Gabor filter can extract robust features against direction and scale changes in the convolutional layer of the residual network, it can better extract image features, and then use the twin neural network model to more accurately identify the similarity of the two images, thereby achieving better recognition effect, so as to more accurately verify the identity of the object to be verified.

[0052] Please refer to Figure 1 , Figure 1 A flowchart of an identity authentication method provided in an embodiment of the present application, the method comprising the following steps:

[0053] Step S110: Acquire a current biometric feature image of the object to be authenticated and acquire a pre-stored historical biometric feature image of the object to be authenticated.

[0054] In the embodiment of the present application, the biometric image may be a fingerprint image, a palm print image, an iris image, a face image or a finger vein image, etc. For ease of description, the finger vein image is used as an example for explanation below. When the biometric image is a finger vein image, the current finger vein image of the object to be verified can be collected, and the object to be verified can be the user to be verified. The electronic device can pre-store historical finger vein images of the object to be verified. For example, when the object to be verified registers an account for the first time, the finger vein image of the object to be verified can be collected, and then stored in association with the account of the object to be verified for subsequent verification.

[0055] When acquiring the historical biometric image, the account information input by the object to be verified can be obtained first, and then the pre-stored historical biometric image corresponding to the account can be searched and obtained according to the account information, while the current biometric image is collected at the current moment. The way to collect the current biometric image or the historical biometric image can be:

[0056] For example, the acquisition device may include a matrix infrared light emitter and a matrix image sensor. Infrared rays may be emitted by the matrix infrared light emitter, and then the finger vein image may be acquired by the matrix image sensor.

[0057] It is understandable that different biometric images have corresponding collection devices that are different.

[0058] The acquisition device can send the acquired finger vein image to the electronic device. Of course, if the finger vein image contains too much background information, the electronic device can also perform preliminary processing on the finger vein image, such as cropping the finger vein image to obtain a finger vein area image, which can facilitate eliminating the influence of background noise on image recognition.

[0059] In some embodiments, the finger vein images collected here may be finger vein images of at least two fingers of the object to be verified. In this case, the historical biometric image includes finger vein images of at least two fingers. Of course, the historical biometric image may include finger vein images of all fingers of the object to be verified.

[0060] For example, the acquisition device can acquire the finger vein images of the index finger, middle finger, and thumb of the object to be verified, and then send them to the electronic device. The electronic device can stitch the finger vein images of these fingers into multiple finger vein images, and the stitching algorithm can be cross-correlation, distance change, structure matching, chain code correlation, etc. In this way, when acquiring the current biometric image, the user to be verified can be prompted to collect finger vein images of multiple fingers, and then the electronic device stitches them into multiple finger vein images before performing subsequent image recognition.

[0061] Step S120: using the image recognition model to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result.

[0062] Among them, the image recognition model is a twin neural network model, and the two network models in the twin neural network model are residual networks including Gabor filters. The two network models are respectively used to extract image features of the current biometric image and the historical biometric image. If one of the network models is used to extract image features of the current biometric image, the other network model is used to extract image features of the historical biometric image.

[0063] The residual network including Gabor filters here can be used to enhance the expression of deep features. It has controllable direction and scaling capabilities, can extract Gabor convolution features in the input image, and then can extract robust features for changes in direction and scale. In addition, the Gabor filter is insensitive to changes in illumination, can improve the adaptability to illumination changes, can tolerate a certain degree of image rotation and deformation, and has a certain degree of robustness to illumination and posture. Therefore, in the application scenario of finger vein image recognition, it can better extract the texture features of the finger vein.

[0064] In addition, the twin neural network model takes two samples as input and outputs its representation embedded in a high-dimensional space to compare the similarity of the two samples. Generally, the twin neural network model includes two network models, which can be neural networks with the same structure, or can be composed of any two neural networks. In the embodiment of the present application, the two network models are residual networks including Gabor filters. The twin neural network model can learn small samples or single samples, and is not easily disturbed by erroneous samples. Therefore, it can be used in recognition scenarios with strict fault tolerance requirements.

[0065] It can be understood that more network models can also be embedded in the twin neural network model, such as three network models, each of which is a residual network containing a Gabor filter. In this case, the input images can be three, that is, two current biometric images of the object to be verified and one historical biometric image can be collected, and then the similarities of the three images can be compared. In this way, more images can be added to participate in the comparison and the recognition accuracy can be improved.

[0066] In the above implementation process, if there are multiple finger vein images to be identified, such as the finger vein images currently acquired include the finger vein images of the index finger, middle finger and thumb of the object to be verified, in this case, its historical biometric image includes the finger vein images of the index finger, middle finger and thumb of the object to be verified (or the historical biometric image can also include the finger vein images of all fingers, and the required image input can be selected when input into the image recognition model). When performing image recognition, the finger vein images of the index finger, middle finger and thumb of the object to be verified currently acquired, and the finger vein images of the index finger, middle finger and thumb of the object to be verified stored historically can be input into the image recognition model, and the image recognition model can compare their similarities in pairs, where the two-by-two comparison refers to comparing the finger vein images of the current index finger with the historical index finger, comparing the finger vein images of the current middle finger with the historical middle finger, and comparing the finger vein images of the current thumb with the historical thumb.

[0067] Alternatively, only two images may be input into the image recognition model at a time, such as first inputting the finger vein image of the current index finger and the finger vein image of the historical index finger into the image recognition model to obtain a similarity, then inputting the finger vein image of the current middle finger and the finger vein image of the historical middle finger into the image recognition model to obtain a similarity, and inputting the finger vein image of the current thumb and the finger vein image of the historical thumb into the image recognition model to obtain a similarity, so that three similarities can be obtained. When performing identity authentication later, the verification is passed when at least one similarity is greater than the set similarity.

[0068] Since the finger vein images of multiple fingers are spliced ​​into multiple finger vein images and input into the image recognition model for recognition, more finger vein information can be included to improve the accuracy of user identity authentication.

[0069] Step S130: verifying the identity of the object to be verified according to the recognition result.

[0070] In some embodiments, the recognition result can be the similarity between the current biometric image and the historical biometric image, so when the similarity is greater than the set similarity, it can be considered that both images belong to the object to be verified, and the identity authentication is passed. If the similarity is less than or equal to the set similarity, it can be considered that the two images belong to different objects, and the identity authentication fails.

[0071] In other embodiments, the recognition result may also be the probability that the current biometric image and the historical biometric image both belong to the object to be verified. If the probability is greater than a set threshold, it is determined that the identity authentication of the object to be verified is passed. If the probability is less than or equal to the set threshold, it is determined that the identity authentication of the object to be verified is not passed.

[0072] The probability here can also be obtained by mapping the similarity between the current biometric image and the historical biometric image. For example, the similarity and the probability can be mapped in advance, and the image recognition model can directly output the corresponding probability according to the similarity.

[0073] However, if the above-mentioned current biometric image and historical biometric images are finger vein images of multiple fingers of the object to be verified, when the image recognition model performs image recognition, as long as the similarity between the finger vein image of at least one finger in the current biometric image and the finger vein image of at least one finger in the historical biometric image is greater than the set similarity, it is considered that the identity authentication of the object to be verified is passed.

[0074] In the above implementation process, an image recognition model is used to identify the similarity between the current biometric image and the historical biometric image of the object to be verified. The image recognition model is a twin neural network model, and the two network models are residual networks including Gabor filters. Since the Gabor filter can extract robust features against direction and scale changes in the convolutional layer of the residual network, it can better extract image features, and then the twin neural network model can more accurately identify the similarity of the two images, thereby achieving better recognition effect, so as to more accurately verify the identity of the object to be verified.

[0075] The following is a brief introduction to the residual network including Gabor filters.

[0076] In order to encode the directional information, the convolution filter in the standard convolutional neural network CNN is modulated by the Gabor filter, and the formula is as follows:

[0077]

[0078] Among them, C i,oIt is a learnable filter, which is a convolution filter through back propagation in the standard CNN; G(u,v) represents a set of Gabor filters with different directions and scales. Represents G(u,v) and C i,o The element-wise product operation between each 2D filter of is the adjusted filter.

[0079] The directional Gabor convolution filter (GOF) is obtained by adjusting the Gabor filter of u direction and v scale. The i-th GOF is represented by the mathematical symbol C i v Indicates, such as:

[0080]

[0081] In GOF, the value of v increases or decreases as the number of layers increases, which means that the scale of the Gabor filter in GOF changes with the number of layers.

[0082] Therefore, the direction information is encoded in the learned filters, and the scale information is embedded in different layers. Due to the direction and scale information contained in the Gabor filters in the GOF, the corresponding convolutional features are enhanced.

[0083] In convolutional neural networks, GOF can be used to generate feature maps that explicitly enhance the scale and orientation information in deep features. The channels of the output feature map are obtained by the following convolution:

[0084]

[0085] Where (n) refers to the input feature maps F and C i The nth channel of i represents the i-th Gabor filter, F represents the input image feature, and F i represents the convolution result of the i-th Gabor filter and F, and u represents F i The k-th direction response.

[0086] The residual network containing Gabor filters can be obtained by deforming tiny ResNet, that is, changing the tiny ResNet convolution to Gabor convolution, and then the residual network containing Gabor filters can be obtained.

[0087] The architecture of Resnet is different from previous architectures in that it can train deeper and deeper networks without reducing performance. However, for the sake of speed and storage space, only an 8-layer network can be retained in the embodiment of the present application, called TinyResNet, and its network depth and width are reduced. TinyResNet divides the entire network into three stages, and the stage width (the number of convolution kernels per layer) is set to 16, 32, and 64 respectively. Figure 2 As shown in (a), the standard convolution is replaced by Gabor convolution to form Figure 2 In the structure shown in (b), Gabor convolution is expressed as GC(u,v), where u and v represent the u direction and v scale.

[0088] On the basis of the above embodiments, the twin neural network model in the embodiments of the present application also includes a concatenation layer, a fully connected layer and a classification layer. When performing image recognition, the image features of the current biometric image can be extracted through one of the two network models to obtain a first feature vector, and the image features of the historical biometric image can be extracted through the other network model to obtain a second feature vector. The first feature vector and the second feature vector are then concatenated through the concatenation layer to obtain a concatenated vector. The concatenated vector is then passed through the fully connected layer and the classification layer in sequence to obtain a recognition result.

[0089] The structural diagram of the twin neural network model in the embodiment of the present application can be shown as follows Figure 3 As shown, the twin neural network model can be trained using a distance function to maximize the distance between non-matching images and minimize the distance between matching images, but the model needs to search for multiple distance functions and different distance thresholds. In addition, even if good results can be obtained through distance measurement, it is still uncertain what the "natural" distance measurement of image space is, and early experiments optimized on the L2 Euclidean distance did not perform well on this dataset. To address this tenderness, you can choose to train the model by outputting a classification layer on two classes (matching and mismatching, matching means the similarity is greater than the set similarity, and mismatching means the similarity is less than or equal to the set similarity).

[0090] The classification layer here can use a softmax layer, so a softmax layer is added to the twin neural network model, allowing the model to learn the representation and distance function that best separates the classes. Of course, the classification layer can also be other structures, such as logistic regression.

[0091] The output of the residual network including the Gabor filter is encoded in an n-dimensional vector. For example, the current biometric image is input to network model 1, and its output is a vector. The historical biometric image is input to network model 2, and its output is b vector. The a vector and the b vector can be concatenated through the concatenation layer to obtain a concatenated vector. The concatenated vector is then input into the fully connected layer, which contains two hidden units to output a feature vector. The output feature vector can be input into the classification layer, and the classification layer outputs the recognition result, such as output probability or similarity.

[0092] Here, when the vectors are concatenated, in order to better extract features, the square of the difference between the first eigenvector and the second eigenvector can be calculated through the concatenation layer, and the Hadamard product between the first eigenvector and the second eigenvector can be calculated through the concatenation layer. Then, the first eigenvector, the second eigenvector, the square of the difference and the Hadamard product can be concatenated through the concatenation layer to obtain a concatenated vector.

[0093] For example, the first eigenvector is represented by a, the second eigenvector is represented by b, and the formula for obtaining the concatenated vector v is as follows:

[0094] v=[ab(ab) 2 a⊙b].

[0095] Of course, the first eigenvector, the second eigenvector and the square of the difference between the two vectors may be connected in series to obtain a series vector, or the first eigenvector, the second eigenvector and the Hadamard product may be connected in series to obtain a series vector.

[0096] In the above implementation process, by connecting two feature vectors, the square of the difference and the Hadamard product in series, more features can be obtained to participate in the recognition, so that the recognition effect is better.

[0097] Based on the above embodiments, the model parameters between the two network models in the twin neural network model can be shared. For example, in the process of training the image recognition model, the model parameters and structure of one of the network models can be set first, and then the model parameters and structure of the network model can be copied to obtain another network model. After the training is completed, if the model parameters of the two network models are not uniform, the model parameters of one of the network models can be copied to the other network model to achieve parameter sharing, which can facilitate the implementation of the network.

[0098] In other implementations, the model parameters between the two network models may not be shared, so that after training is completed, the two network models each have different model parameters, which is conducive to extracting differentiated features in the input image.

[0099] Taking the use of finger vein images for identity authentication as an example, the effects of the embodiments of the present application are illustrated below through some experimental data.

[0100] 1. Dataset.

[0101] The experiments use two widely used public finger vein datasets: FV-USM and MMCBNU_6000.

[0102] FV-USM consists of 2952 JPG images with a resolution of 640*480 from 123 volunteers, each with 4 fingers, each finger is a different category, so there are 492 categories in total. The images were collected in two different periods, with 6 images per finger in each period. The dataset provides the region of interest (ROI) extracted from the original image with a resolution of 50*150. In the following experiments, the ROI image extracted from period 1 is used directly.

[0103] The MMCBNU_6000 dataset is another public finger vein database created by 100 volunteers from 20 countries. Each person provided ten image samples from the index finger, middle finger, and ring finger of both hands. Therefore, MMCBNU_6000 contains 600 classes, 6000 images, each of which is a grayscale image stored in BMP format with a resolution of 640*480. In addition, the dataset also provides ROI extracted from the original image with a resolution of 128*60.

[0104] See Table 1 for details. Figure 4 Some sample images of the two datasets are given. It should be noted that Figure 4 (a),In FV-USM, the original direction of the fingertip is downward, e.g. Figure 4 (b), while in MMCBNU_6000, the original direction of the fingertip is to the right. In order to verify the scalability of the proposed network orientation, the original fingertip orientation is maintained in the experiment instead of rotating it to the same orientation.

[0105] Table 1

[0106]

[0107] 2. Experimental preparation.

[0108] There are two types of experiments: open set experiments and closed set experiments. In closed set experiments, all available fingers participate in training and testing. For practical applications, closed set scenarios are unreasonable. On the contrary, in open set experiments, fingers are divided into two categories, participating in training and testing respectively, and the training set and test set are disjoint. Obviously, open set experiments are more in line with the actual situation, so open set experiments are performed in the following experiments.

[0109] Each dataset was randomly split into a training set and a test set according to the approximate ratio of 90% and 10%, and 10-fold cross validation was performed to report the experimental results.

[0110] Therefore, for FV-USM, there are 50 classes (300 images) for testing and the remaining 442 classes (2652 images) for training. For MMCBNU_6000, there are 60 classes (600 images) for testing and the remaining 540 classes (5400 images) for training.

[0111] The image recognition model of this application needs to generate image pairs. For positive matching, a finger is sampled and then other different images captured from the same finger are traversed. For negative matching, five images are randomly selected from other categories and a finger is sampled. The ratio of positive to negative examples is 1:5, as shown in Table 2.

[0112] Table 2

[0113]

[0114] The experiment uses FAR (false acceptance rate), FRR (false rejection rate), and EER (equal error rate) as evaluation criteria. FAR is the ratio of the number of false acceptances divided by the number of authentication attempts. FRR is the ratio of the number of false rejections divided by the number of authentication attempts. EER is the rate at which FAR equals FRR. The lower the EER, the better the algorithm. EER refers to the most common evaluation protocol for vein authentication. In addition to EER, verification accuracy (ACC) can also be applied for evaluation.

[0115] 3. Parameter selection.

[0116] (1) Scale and orientation selection: First, the number of scales and orientations is tested to select the parameters with the best performance at high ACC and low EER. Tables 3 and 4 below give the ACC, EER, and time / batch for different scales and orientations on FV-USM and MMCBNU_6000, respectively. The numbers 1 to 6 in the first 6 columns of the table represent all Gabor convolution layers with a fixed scale, and the last column "Increase" in the table represents each Gabor convolution layer with increasing scale from 1 to 5. The results show that the results get better when the scale increases, but when the scale reaches a certain peak, the results begin to deteriorate. The results show that Gabor convolution at a fixed scale is better than Gabor convolution at increasing scale. The first row of the 4 directions is the validation accuracy (ACC), and the second row is the EER. The third row of the 4 directions is the computation time for training a batch of different parameters, and the row of 8 directions has the same architecture as the row of 4 directions. The batch size is set to 32 and the learning rate is set to 0.00002. The number of training rounds is 10. Table 3 shows the results of FV-USM. The parameter combination of 4 directions and 5 scales performs slightly better than the parameter combination of 4 directions and 4 scales on ACC and EER, while the batch training time of 5 scales is longer than that of 4 scales. Table 4 shows the results on MMCBNU_6000. The parameter combination of 4 directions and 4 scales performs better than other combinations on the evaluation criteria. Figure 5 (a) and (b) show that in different datasets, with the same direction parameters, the training time of a batch (batch size is 32) fluctuates slightly between different scales, and it is obvious that the training time for 4 directions is significantly smaller than that for 8 directions.

[0117] Table 3

[0118]

[0119] Table 4

[0120]

[0121] After comprehensive consideration, the parameters of 4 direction-4 scale combinations are selected in the following experiments (from Tables 3 and 4).

[0122] (2) Output dimension selection: The output of the image recognition model in this application is encoded as an n-dimensional vector, and different n are tested between 100 and 300, and the results are shown in Tables 5 and 6. When n is 200, the performance is better.

[0123] For example, the output of input image A after being processed by one of the network models is recorded as vector a, and the output of input image B after being processed by the other network model is recorded as vector b. These vectors, their squared differences, and the Hadamard product of each vector are concatenated. This will result in the following vector of size 4n:

[0124] v=[ab(ab) 2 a⊙b].

[0125] Different n values ​​vary from 100 to 300, and the performance is better when n value is 200.

[0126] Table 5 Effect of different output sizes on experimental results

[0127]

[0128]

[0129] Table 6 Effect of different output cascades on experimental results

[0130]

[0131] (3) The distance company comparison is shown in Table 7:

[0132] Table 7 Acc and EER calculated using different similarity metrics on FV-USM

[0133]

[0134] 4. Compare with other networks.

[0135] The image recognition model of this application is compared with other recently developed neural networks, such as GCN, Densenet161, VGG16 Gabor, Resnet18 Gabor, Alexnet Gabor. The input sizes of these networks are different. As far as general knowledge is concerned, the larger the input size, the better the results are at the expense of model size. For fair comparison, the input size of these networks is modified to 105*105, and all are untrained. For generality, the experiments are conducted on two databases, FV-USM and MMCBNU_6000.

[0136] Table 8 shows the results of FV-USM. The image recognition model of this application not only has better performance with an accuracy of 99.61% and an EER of 0.46%, but also has a smaller model size of 7.71M.

[0137] Table 9 shows the results on MMCBNU_6000. The accuracy of Densenet18 is 99.85% and the EER is 0.16%, which is slightly better than the image recognition model of this application with an accuracy of 99.41% and an EER of 0.99%. However, the model size of Densenet18 is 219 meters, which is much larger than the image recognition model of this application.

[0138] Table 8

[0139]

[0140] Table 9

[0141]

[0142] like Figure 6 The left column of (a) and (b) shows the relationship between model size and ACC under different datasets (the image recognition model of this application is Figure 6 denoted as SGRN in Figure 6 The right column of (a) and (b) shows the relationship between model size and EER on all databases. The SGRN model of this application is smaller in size, has higher ACC and lower EER.

[0143] Figure 7 The comparison of DET (detection error trade-off) curves of different methods on different finger vein image sets is shown, as Figure 7 (a) is a schematic diagram of the relationship between the size of each model and DET under the FV-USM dataset. Figure 7 (b) is a schematic diagram of the relationship between the size of each model and DET under the MMCBNU_6000 dataset. In general, the SGRM (the image recognition model of this application) of this application is Figure 7 The performance of SGRN in Fig.

[0144] Therefore, through the above experiments, it can be seen that the image recognition model proposed in this application can directly realize end-to-end verification. Compared with other types of models, it has higher verification performance, smaller model size, occupies less storage resources and less computing resources, and is more conducive to deployment in low-computing power computing devices.

[0145] Please refer to Figure 8 , Figure 8 This is a structural block diagram of an identity verification device 200 provided in an embodiment of the present application. The device 200 may be a module, program segment or code on an electronic device. It should be understood that the device 200 is similar to the above Figure 1 The method embodiment corresponds to and can be executed Figure 1The various steps involved in the method embodiment and the specific functions of the device 200 can be found in the above description. To avoid repetition, the detailed description is appropriately omitted here.

[0146] Optionally, the device 200 includes:

[0147] An image acquisition module 210 is used to acquire a current biometric image of the object to be authenticated and to acquire a pre-stored historical biometric image of the object to be authenticated;

[0148] The recognition module 220 is used to use an image recognition model to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result, wherein the image recognition model is a twin neural network model, and two network models in the twin neural network model are residual networks including Gabor filters, and the two network models are respectively used to extract image features of the current biometric image and the historical biometric image;

[0149] The verification module 230 is used to verify the identity of the object to be verified according to the recognition result.

[0150] Optionally, the twin neural network model also includes a concatenation layer, a fully connected layer and a classification layer, and the recognition module 220 is used to extract the image features of the current biometric image through one of the two network models to obtain a first feature vector, and to extract the image features of the historical biometric image through another network model to obtain a second feature vector; concatenate the first feature vector and the second feature vector through the concatenation layer to obtain a concatenation vector; and pass the concatenation vector through the fully connected layer and the classification layer in sequence to obtain a recognition result.

[0151] Optionally, the identification module 220 is used to calculate the square of the difference between the first eigenvector and the second eigenvector through the concatenation layer, and to calculate the Hadamard product between the first eigenvector and the second eigenvector through the concatenation layer; and to concatenate the first eigenvector, the second eigenvector, the square of the difference and the Hadamard product through the concatenation layer to obtain a concatenated vector.

[0152] Optionally, the recognition result is the probability that the current biometric image and the historical biometric image both belong to the object to be verified, and the verification module 230 is used to determine that the identity authentication of the object to be verified is passed if the probability is greater than a set threshold; if the probability is less than or equal to the set threshold, determine that the identity authentication of the object to be verified is not passed.

[0153] Optionally, model parameters are shared between the two network models.

[0154] Optionally, the biometric image is a fingerprint image, a palm print image, an iris image or a finger vein image.

[0155] Optionally, when the biometric image is a finger vein image, the image acquisition module 210 is used to acquire current finger vein images of at least two fingers of the object to be authenticated, wherein the historical biometric image includes historical finger vein images of the at least two fingers.

[0156] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0157] Please refer to Fig. 9 , Fig. 9 A structural diagram of an electronic device for executing an identity authentication method provided in an embodiment of the present application, the electronic device may include: at least one processor 310, such as a CPU, at least one communication interface 320, at least one memory 330 and at least one communication bus 340. Among them, the communication bus 340 is used to realize direct connection and communication between these components. Among them, the communication interface 320 of the device in the embodiment of the present application is used to communicate signaling or data with other node devices. The memory 330 can be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 330 can optionally also be at least one storage device located away from the aforementioned processor. Computer-readable instructions are stored in the memory 330. When the computer-readable instructions are executed by the processor 310, the electronic device executes the above-mentioned Figure 1 The method process is shown.

[0158] Understandably, Fig. 9 The structure shown is for illustration only, and the electronic device may also include Fig. 9 More or fewer components as shown, or with Fig. 9 Different configurations are shown. Fig. 9 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0159] The present application embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the following is performed: Figure 1 The method process in the method embodiment shown is executed by the electronic device.

[0160] The present embodiment discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments, for example, including: obtaining a current biometric image of an object to be verified and obtaining a pre-stored historical biometric image of the object to be verified; using an image recognition model to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result, wherein the image recognition model is a twin neural network model, and two network models in the twin neural network model are residual networks including Gabor filters, and the two network models are respectively used to extract image features of the current biometric image and the historical biometric image; and verifying the identity of the object to be verified according to the recognition result.

[0161] In summary, the embodiments of the present application provide an identity authentication method, device, electronic device and readable storage medium. The method uses an image recognition model to identify the similarity between the current biometric image and the historical biometric image of the object to be verified. The image recognition model is a twin neural network model, and the two network models are residual networks including Gabor filters. Since the Gabor filter can extract robust features against direction and scale changes in the convolutional layer of the residual network, it can better extract image features, and then use the twin neural network model to more accurately identify the similarity of the two images, thereby achieving better recognition effect, so as to more accurately verify the identity of the object to be verified.

[0162] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0163] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0164] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0165] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0166] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An identity authentication method, characterized in that: The method comprises: Acquire a current biometric image of the object to be authenticated and acquire a pre-stored historical biometric image of the object to be authenticated; Using an image recognition model to identify the similarity between the current biometric image and the historical biometric image, and obtaining a recognition result, wherein the image recognition model is a twin neural network model, two network models in the twin neural network model are residual networks including Gabor filters, and the two network models are respectively used to extract image features of the current biometric image and the historical biometric image; Verifying the identity of the object to be verified according to the recognition result; The twin neural network model further includes a concatenation layer, a fully connected layer, and a classification layer. The image recognition model is used to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result, including: Extracting image features of the current biometric image through one of the two network models to obtain a first feature vector, and extracting image features of the historical biometric image through another network model to obtain a second feature vector; Concatenate the first feature vector and the second feature vector through the concatenation layer to obtain a concatenated vector; Passing the series vector through the fully connected layer and the classification layer in sequence to obtain a recognition result; The step of concatenating the first feature vector and the second feature vector through the concatenation layer to obtain a concatenated vector includes: calculating the square of the difference between the first eigenvector and the second eigenvector through the concatenated layer, and calculating the Hadamard product between the first eigenvector and the second eigenvector through the concatenated layer; The first eigenvector, the second eigenvector, the square of the difference, and the Hadamard product are connected in series through the connection layer to obtain a series connection vector.

2. The method according to claim 1, characterized in that The recognition result is a probability that both the current biometric image and the historical biometric image belong to the object to be verified, and verifying the identity of the object to be verified according to the recognition result includes: If the probability is greater than a set threshold, it is determined that the identity authentication of the object to be authenticated is passed; If the probability is less than or equal to the set threshold, it is determined that the identity authentication of the object to be authenticated has failed.

3. The method according to claim 1, characterized in that Model parameters are shared between the two network models.

4. The method according to any one of claims 1 to 3, characterized in that: The biometric image is a fingerprint image, a palm print image, an iris image or a finger vein image.

5. The method according to any one of claims 1 to 3, characterized in that: When the biometric image is a finger vein image, the step of obtaining the current biometric image of the object to be verified includes: Current finger vein images of at least two fingers of the object to be authenticated are acquired, wherein the historical biometric image includes the historical finger vein images of the at least two fingers.

6. An identity verification device, characterized in that: The device comprises: An image acquisition module, used to acquire a current biometric image of the object to be authenticated and to acquire a pre-stored historical biometric image of the object to be authenticated; A recognition module, used to use an image recognition model to identify the similarity between the current biometric image and the historical biometric image to obtain a recognition result, wherein the image recognition model is a twin neural network model, two network models in the twin neural network model are residual networks including Gabor filters, and the two network models are respectively used to extract image features of the current biometric image and the historical biometric image; A verification module, used to verify the identity of the object to be verified according to the recognition result; The twin neural network model further includes a concatenation layer, a fully connected layer and a classification layer, and the recognition module is specifically used to extract the image features of the current biometric image through one of the two network models to obtain a first feature vector, and to extract the image features of the historical biometric image through another network model to obtain a second feature vector; concatenate the first feature vector and the second feature vector through the concatenation layer to obtain a concatenation vector; and pass the concatenation vector through the fully connected layer and the classification layer in sequence to obtain a recognition result; Among them, the recognition module is specifically used to calculate the square of the difference between the first eigenvector and the second eigenvector through the concatenation layer, and to calculate the Hadamard product between the first eigenvector and the second eigenvector through the concatenation layer; the first eigenvector, the second eigenvector, the square of the difference and the Hadamard product are concatenated through the concatenation layer to obtain a concatenated vector.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 5 is executed.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is performed.

Images