A biometric feature recognition method, apparatus, device and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for biometric recognition. The method includes: acquiring four-dimensional tensor data based on biological images, which contains information on the number of biological images, the number of input channels, and the width and height of the images; inputting the four-dimensional tensor data into a biometric recognition model; and outputting recognition features with the same number of biological images as the number of images. The dimension of each recognition feature is greater than the number of input channels. The main structure of the biometric recognition model is a MobileNetV2 structure, and the activation function used in the main structure is the FReLU activation function. It is evident that because the biometric recognition model introduces a feature recognition module, it can align biometric features with corresponding positions of biological poses. Furthermore, the FReLU activation function in the main structure is a linear function, which can reduce the risk of overfitting during model training and improve the accuracy of biometric recognition.

Description

Technical Field

[0001] This application relates to the field of identification technology, and more specifically, to a method, apparatus, device, and storage medium for identifying biometric features. Background Technology

[0002] With the continuous development of information technology, information mining has attracted the interest of many researchers, who aim to extract useful information from massive amounts of data. Different objects possess unique characteristics, such as fingerprints and palm prints. Researchers can extract feature information from complex images of fingers or palms to identify a specific object. For image feature extraction methods, researchers can design algorithms or train specialized models to perform feature extraction from images. Typically, before inputting an image into an algorithm / model, it needs to be processed to conform to the algorithm / model's data input format, such as the format of four-dimensional tensor data.

[0003] In current biometric recognition models, changes in biological posture have a significant impact on the recognized biometric features. Furthermore, the activation functions in these models introduce numerous nonlinear transformations, leading to a high risk of overfitting during training and consequently, low accuracy in biometric recognition. Summary of the Invention

[0004] In view of the above problems, this application is made to provide a method, apparatus, device and storage medium for biometric identification, so as to reduce the risk of training overfitting and improve the accuracy of biometric identification.

[0005] To achieve the above objectives, the following specific solutions are proposed:

[0006] A biometric identification method, comprising:

[0007] Obtain four-dimensional tensor data based on biological images. The initial four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height.

[0008] The four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features with the same number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

[0009] Optionally, the main structure of the model includes downsampling convolutional blocks;

[0010] The four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features equal in number to the number of biological images, including:

[0011] By using the downsampling convolution block in the main structure of the pre-established biometric recognition model, the four-dimensional tensor data is downsampled five times to obtain four-dimensional tensor data to be recognized with the same number of images as the number of biological images. The number of channels in the four-dimensional tensor data to be recognized is greater than the number of input channels.

[0012] The feature recognition module of the biometric recognition model performs feature recognition on the four-dimensional tensor data to be recognized, and obtains the same number of recognition features as the number of biological images. The number of dimensions of each recognition feature is greater than the number of channels of the four-dimensional tensor data to be recognized.

[0013] Optionally, the image width and image height of the four-dimensional tensor data are both 256 pixels, and the number of input channels of the four-dimensional tensor data is 3;

[0014] The four-dimensional tensor data is downsampled five times using downsampling convolutional blocks in the main structure of a pre-established biometric recognition model to obtain the four-dimensional tensor data to be recognized, including:

[0015] The four-dimensional tensor data is downsampled for the first time by using the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain the first four-dimensional tensor data. The image width and image height of the first four-dimensional tensor data are both 128 pixels.

[0016] The first four-dimensional tensor data is downsampled in the second round by the downsampling convolution block to obtain the second four-dimensional tensor data. The image width and image height of the second four-dimensional tensor data are both 64 pixels.

[0017] The third round of downsampling processing is performed on the second four-dimensional tensor data by the downsampling convolution block to obtain the third four-dimensional tensor data. The image width and image height of the third four-dimensional tensor data are both 32 pixels.

[0018] The fourth round of downsampling processing is performed on the third four-dimensional tensor data by the downsampling convolution block to obtain the fourth four-dimensional tensor data. The image width and image height of the fourth four-dimensional tensor data are both 16 pixels.

[0019] The fourth four-dimensional tensor data is downsampled for the fifth time by the downsampling convolution block to obtain the four-dimensional tensor data to be identified. The image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384.

[0020] Optionally, the image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384.

[0021] The feature recognition module of the biometric recognition model performs feature recognition on the four-dimensional tensor data to be identified, obtaining the same number of recognition features as the biological image, including:

[0022] Following the processing order of convolution, normalization, and max pooling, the feature recognition module of the biometric recognition model processes the four-dimensional tensor data to be identified in two rounds to obtain intermediate four-dimensional tensor data. The image width and image height of the intermediate four-dimensional tensor data are both 2 pixels.

[0023] The feature recognition module performs channel transformation convolution, normalization and max pooling on the intermediate four-dimensional tensor data to obtain the target four-dimensional tensor data. The image width and image height of the target four-dimensional tensor data are both single pixels, and the number of dimensions of the target four-dimensional tensor data is 512.

[0024] The feature recognition module reshapes the target four-dimensional tensor data to obtain the same number of recognition features as the biological image.

[0025] A biometric identification device, comprising:

[0026] Tensor data acquisition unit is used to acquire four-dimensional tensor data based on biological images. The initial four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height.

[0027] The feature recognition unit is used to input the four-dimensional tensor data into a pre-established biometric recognition model and output the same number of recognition features as the number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

[0028] Optionally, the main structure of the model includes downsampling convolutional blocks;

[0029] The feature recognition unit includes:

[0030] The main structure recognition unit is used to perform five rounds of downsampling processing on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain four-dimensional tensor data to be recognized with the same number of images as the number of biological images. The number of channels in the four-dimensional tensor data to be recognized is greater than the number of input channels.

[0031] The identification module identification unit is used to perform feature identification on the four-dimensional tensor data to be identified through the feature identification module of the biometric identification model, and obtain identification features with the same number as the number of biological images, wherein the number of dimensions of each identification feature is greater than the number of channels of the four-dimensional tensor data to be identified.

[0032] Optionally, the image width and image height of the four-dimensional tensor data are both 256 pixels, and the number of input channels of the four-dimensional tensor data is 3;

[0033] The main structure identification unit includes:

[0034] The first main structure recognition subunit is used to perform the first round of downsampling processing on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain the first four-dimensional tensor data. The image width and image height of the first four-dimensional tensor data are both 128 pixels.

[0035] The second main structure recognition subunit is used to perform a second round of downsampling processing on the first four-dimensional tensor data through the downsampling convolution block to obtain the second four-dimensional tensor data. The image width and image height of the second four-dimensional tensor data are both 64 pixels.

[0036] The third main structure recognition subunit is used to perform a third round of downsampling processing on the second four-dimensional tensor data through the downsampling convolution block to obtain the third four-dimensional tensor data. The image width and image height of the third four-dimensional tensor data are both 32 pixels.

[0037] The fourth main structure recognition subunit is used to perform a fourth round of downsampling processing on the third four-dimensional tensor data through the downsampling convolution block to obtain the fourth four-dimensional tensor data, wherein the image width and image height of the fourth four-dimensional tensor data are both 16 pixels.

[0038] The fifth main structure recognition subunit is used to perform a fifth round of downsampling processing on the fourth four-dimensional tensor data through the downsampling convolution block to obtain the four-dimensional tensor data to be recognized. The image width and image height of the four-dimensional tensor data to be recognized are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be recognized is 384.

[0039] Optionally, the image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384.

[0040] The identification module identification unit includes:

[0041] The first recognition module recognition subunit is used to process the four-dimensional tensor data to be recognized in two rounds through the feature recognition module of the biometric recognition model according to the processing order of first convolution, then normalization, and finally maximum pooling, to obtain intermediate four-dimensional tensor data. The image width and image height of the intermediate four-dimensional tensor data are both 2 pixels.

[0042] The second recognition module recognition subunit is used to perform channel transformation convolution, normalization processing and max pooling processing on the intermediate four-dimensional tensor data through the feature recognition module to obtain target four-dimensional tensor data. The image width and image height of the target four-dimensional tensor data are both single pixels, and the number of dimensions of the target four-dimensional tensor data is 512.

[0043] The third identification module identification subunit is used to reshape the target four-dimensional tensor data through the feature identification module to obtain the same number of identification features as the biological image.

[0044] A biometric identification device, including a memory and a processor;

[0045] The memory is used to store programs;

[0046] The processor is used to execute the program to implement the various steps of the biometric identification method as described above.

[0047] A storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the biometric identification method as described above.

[0048] Using the above technical solution, this application obtains four-dimensional tensor data based on biological images. The initial four-dimensional tensor data contains information about the number of biological images, the number of input channels, the image width, and the image height. Further, the four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features with the same number of biological images. The number of dimensions of each recognition feature is equal to the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure and includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer. Therefore, it can be seen that since the biometric recognition model introduces a feature recognition module, it can align the biometric features corresponding to the biometric pose. Since the FReLU activation function in the main structure of the model is a linear function, it can reduce the risk of overfitting during the training process and improve the recognition accuracy of biometric features. Attached Figure Description

[0049] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0050] Figure 1 A schematic diagram of a process for identifying biometric features provided in an embodiment of this application;

[0051] Figure 2 This application provides a schematic diagram of a downsampling process for obtaining four-dimensional tensor data to be identified, as illustrated in an embodiment of the present application.

[0052] Figure 3 This is a schematic diagram illustrating the feature recognition process of a feature recognition module provided in an embodiment of this application.

[0053] Figure 4 A schematic diagram of a device for identifying biometric features provided in an embodiment of this application;

[0054] Figure 5 This is a schematic diagram of the structure of a biometric identification device provided in an embodiment of this application. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] The proposed solution can be implemented based on a terminal with data processing capabilities, such as a computer, server, or cloud platform.

[0057] Next, combined Figure 1 The biometric identification method of this application may include the following steps:

[0058] Step S110: Obtain four-dimensional tensor data based on biological images.

[0059] Among them, the four-dimensional tensor data based on biological images can contain information about the number of biological images, the number of input channels, the image width, and the image height.

[0060] Specifically, the format of four-dimensional tensor data can be represented as [A, B, C, D]. A represents the number of biological images, which can indicate the batch size of the biological images or the number of biological images required to input the four-dimensional tensor data into the biometric recognition model. Biological images can be images containing biological information, such as palm or finger images, and can be obtained from existing biological information databases. B represents the number of input channels, indicating the number of channels through which the four-dimensional tensor data is input into the biometric recognition model. Each channel corresponds to a feature dimension of the biological image. C represents the image width, indicating the pixel value of the biological image in the width direction. D represents the image height, indicating the pixel value of the biological image in the height direction.

[0061] It is understandable that four-dimensional tensor data can be generated based on an image A with a width of C and a height of D, and the input channel parameter B of the biometric recognition model, so that the four-dimensional tensor data conforms to the input format of the biometric recognition model.

[0062] Step S120: Input the four-dimensional tensor data into the pre-established biometric recognition model and output the same number of recognition features as the number of biological images.

[0063] Specifically, the number of dimensions for each recognition feature can be the number of output channels of the biometric recognition model.

[0064] The number of output channels can be greater than the number of input channels. The biometric recognition model can consist of a main model structure and a feature recognition module. The main model structure can be a MobileNetV2 structure, and the activation function used in the main model structure can be the FReLU activation function. The feature recognition module can be attached to the end of the main model structure and can include a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

[0065] Understandably, the feature recognition module can use the convolution function of the convolutional layer to expand the feature dimension of the four-dimensional tensor data image from the number of input channels to the number of output channels. Therefore, the feature recognition module can accept four-dimensional tensor data with the same number of input channels as input and four-dimensional tensor data with the same number of output channels as output, providing richer image feature information and richer empirical values ​​for the feature recognition module.

[0066] The biometric recognition method provided in this embodiment acquires four-dimensional tensor data based on biological images. The initial four-dimensional tensor data contains information about the number of biological images, the number of input channels, the image width, and the image height. Further, the four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features with the same number of biological images. The number of dimensions of each recognition feature is equal to the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure and includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer. Therefore, it can be seen that since the biometric recognition model introduces a feature recognition module, it can align the biometric features corresponding to the biometric pose. Since the FReLU activation function in the main structure of the model is a linear function, it can reduce the risk of overfitting during the training process and improve the recognition accuracy of biometric features.

[0067] In some embodiments of this application, the main structure of the biometric recognition model mentioned in the above embodiments is described. The main structure of the model may include a downsampling convolution block, which can downsample the input four-dimensional tensor data. The stride parameter of the downsampling convolution block has a value of 2.

[0068] Based on this, the process of inputting four-dimensional tensor data into a pre-established biometric recognition model and outputting recognition features equal in number to the biological images is described above. This process may include:

[0069] S1. Using the downsampling convolution block in the main structure of the pre-established biometric recognition model, the four-dimensional tensor data is downsampled five times to obtain the same number of four-dimensional tensor data to be recognized as the number of biological images.

[0070] Specifically, the number of channels in the four-dimensional tensor data to be identified can be greater than the number of input channels.

[0071] Understandably, since the stride parameter of the downsampling convolution block is 2, the width and height of the resulting four-dimensional tensor image are halved after each downsampling process compared to the original four-dimensional tensor image. Therefore, after five rounds of downsampling, the width and height of the four-dimensional tensor image to be recognized is 1 / 32 of the original four-dimensional tensor image. Because the width and height of the four-dimensional tensor image to be recognized is reduced, its feature dimensions increase; therefore, the number of channels in the four-dimensional tensor image to be recognized can be greater than the number of input channels.

[0072] For example, when the image width of the four-dimensional tensor data input to the biometric recognition model is 256 pixels and its image height is 256 pixels, after 5 rounds of downsampling processing, the resulting image width of the four-dimensional tensor data to be recognized is 8 pixels and the image height is 8 pixels.

[0073] S2. The feature recognition module of the biometric recognition model performs feature recognition on the four-dimensional tensor data to be recognized, and obtains the same number of recognition features as the biological image.

[0074] Specifically, the feature recognition module can accept several batches of biological image inputs and output the same number of recognition features.

[0075] In this case, the number of dimensions of each identification feature can be greater than the number of channels of the four-dimensional tensor data to be identified.

[0076] It is understandable that, since the feature recognition module can expand the feature dimensions of the four-dimensional tensor data image through the convolution function of the convolution layer, the number of dimensions of the recognized features can be greater than the number of channels of the four-dimensional tensor data to be recognized.

[0077] For example, if the number of channels in the four-dimensional tensor data to be identified is 384, then after the feature recognition module performs feature recognition on the four-dimensional tensor data to be identified, the number of dimensions of the identified features is 512. That is, the feature recognition module's recognition result for the four-dimensional tensor data to be identified is 512 dimensional features.

[0078] In some embodiments of this application, taking an image width of 256 pixels, an image height of 256 pixels, and 3 input channels as an example, the process of performing 5 rounds of downsampling on the four-dimensional tensor data to obtain the same number of four-dimensional tensor data to be identified as the number of biological images is described, using the downsampling convolution block in the main structure of the pre-established biometric recognition model mentioned in the above embodiments. Figure 2 As shown, the process may include:

[0079] Step S210: Perform the first round of downsampling on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain the first four-dimensional tensor data.

[0080] It is understandable that, since the stride parameter of the downsampling convolution block is 2, the width and height of the resulting four-dimensional tensor data image are half the size of the original four-dimensional tensor data image after each downsampling process. Therefore, the width and height of the first four-dimensional tensor data image are both 128 pixels.

[0081] Step S220: Perform a second round of downsampling on the first four-dimensional tensor data using downsampling convolutional blocks to obtain the second four-dimensional tensor data.

[0082] It is understandable that, since the stride parameter of the downsampling convolution block is 2, the width and height of the resulting four-dimensional tensor data image are half the size of the original four-dimensional tensor data image after each downsampling process. Therefore, the width and height of the second four-dimensional tensor data image are both 64 pixels.

[0083] Step S230: Perform a third round of downsampling on the second four-dimensional tensor data using downsampling convolutional blocks to obtain the third four-dimensional tensor data.

[0084] It is understandable that, since the stride parameter of the downsampling convolution block is 2, the width and height of the resulting four-dimensional tensor data image are half the size of the original four-dimensional tensor data image after each downsampling process. Therefore, the width and height of the third four-dimensional tensor data image are both 32 pixels.

[0085] Step S240: Perform a fourth round of downsampling on the third four-dimensional tensor data using downsampling convolutional blocks to obtain the fourth four-dimensional tensor data.

[0086] It is understandable that, since the stride parameter of the downsampling convolution block is 2, the width and height of the resulting four-dimensional tensor data image are half the size of the original four-dimensional tensor data image after each downsampling process. Therefore, the width and height of the fourth four-dimensional tensor data image are both 16 pixels.

[0087] Step S250: Perform a fifth round of downsampling processing on the fourth four-dimensional tensor data using downsampling convolutional blocks to obtain the four-dimensional tensor data to be identified.

[0088] Understandably, since the stride parameter of the downsampling convolution block is 2, the width and height of the resulting four-dimensional tensor image are half the size of the original four-dimensional tensor image after each downsampling process. Therefore, the width and height of the four-dimensional tensor image to be identified are both 8 pixels. After 5 rounds of downsampling, the number of channels in the four-dimensional tensor data to be identified increases from 3 channels in the initial input four-dimensional tensor data to 384 channels.

[0089] The biometric identification method provided in this embodiment performs five rounds of downsampling processing on the four-dimensional tensor data through downsampling convolutional blocks to obtain four-dimensional tensor data to be identified with smaller pixels and more channels. Since the pixel size of the four-dimensional tensor data to be identified is smaller, it is easier to store and use. Since the number of channels of the four-dimensional tensor data to be identified is greater, the feature dimensions of the image are richer, providing richer empirical values ​​for the biometric identification model during the training process.

[0090] In some embodiments of this application, taking an image width of 8 pixels, an image height of 8 pixels, and 384 input channels as examples, the process of using the feature recognition module of the biometric recognition model mentioned in the above embodiments to perform feature recognition on the four-dimensional tensor data to be recognized, and obtaining the same number of recognition features as the biological image, is described. Figure 3 As shown, the process may include:

[0091] Step S310: Following the processing order of convolution, normalization, and max pooling, the feature recognition module of the biometric recognition model processes the four-dimensional tensor data to be recognized in two rounds to obtain the intermediate four-dimensional tensor data.

[0092] Understandably, since the stride parameter in the pooling process is 2, the image size of the obtained four-dimensional tensor data is halved after each round of processing (convolution → normalization → max pooling). Therefore, after two rounds of processing, the image width and image height of the obtained intermediate four-dimensional tensor data are both 2 pixels.

[0093] The number of channels in the intermediate four-dimensional tensor data is 384, which is the number of input channels of the four-dimensional tensor data to be identified, while the number of images remains unchanged.

[0094] Specifically, following the processing order of convolution, normalization, and finally max pooling, the process of processing the four-dimensional tensor data to be identified through the biometric recognition model in two rounds to obtain the intermediate four-dimensional tensor data can include:

[0095] S3101. The first convolutional layer in the feature recognition module performs convolution processing on the four-dimensional tensor data to be recognized to obtain the first data.

[0096] Specifically, the kernel size of the first convolutional layer is 3*3. The first convolutional layer is used to receive input of 384-channel four-dimensional tensor data and output 384-channel data. The padding parameter is 1.

[0097] S3102. The first data is normalized through the first batch normalization layer in the feature recognition module to obtain the second data.

[0098] S3103. The second data is processed by the first maximum pooling layer in the feature recognition module to obtain the third data, thus completing one round of processing of the four-dimensional tensor data to be recognized.

[0099] Specifically, the pooling kernel size of the first maximum pooling layer is 2*2, and the stride parameter is 2.

[0100] It is understandable that after the four-dimensional tensor data to be identified is processed in sequence through the first convolutional layer, the first batch normalization layer, and the first max pooling layer, the width and height of the resulting third data image are halved. Therefore, the width and height of the third data image are both 4 pixels.

[0101] S3104. The third data is processed by convolution through the first convolutional layer to obtain the fourth data.

[0102] S3105. The fourth data is processed by convolution through the first batch normalization layer to obtain the fifth data.

[0103] S3106. The fifth data is processed by max pooling through the first max pooling layer to obtain the intermediate four-dimensional tensor data.

[0104] It is understandable that S3104-S3106 is a repeat of step S3101-S3103. Therefore, the width and height of the resulting intermediate four-dimensional tensor data image are half the width and height of the third data image. Thus, the width and height of the intermediate four-dimensional tensor data image are both 2 pixels.

[0105] The intermediate four-dimensional tensor data obtained after two rounds of processing has 384 channels.

[0106] Step S320: The intermediate four-dimensional tensor data is processed by channel transformation convolution, normalization and max pooling through the feature recognition module to obtain the target four-dimensional tensor data.

[0107] Understandably, since the target four-dimensional tensor data is obtained by processing the intermediate four-dimensional tensor data through one round of (convolution → normalization → max pooling), the image width and height of the target four-dimensional tensor data are both single pixels. This round of processing can utilize the second convolutional layer, the second batch normalization layer, and the second max pooling layer.

[0108] In this round of processing, the number of dimensional features of the target four-dimensional tensor data was increased; therefore, the number of dimensions of the target four-dimensional tensor data is equal to the number of output channels. The number of images in the target four-dimensional tensor data remains unchanged.

[0109] Specifically, the process of obtaining the target four-dimensional tensor data by performing channel transformation convolution, normalization, and max pooling on the intermediate four-dimensional tensor data through the feature recognition module can include:

[0110] S3201. The second convolutional layer in the feature recognition module performs channel transformation convolution on the intermediate four-dimensional tensor data to obtain the sixth data.

[0111] The first convolutional layer has a kernel size of 3*3 and a padding parameter of 1. It can receive 384 channels of four-dimensional tensor data as input and output 512 channels of data.

[0112] S3202. The intermediate four-dimensional tensor data is normalized through the second batch normalization layer in the feature recognition module to obtain the seventh data.

[0113] S3203. The seventh data is processed by the second maximum pooling layer in the feature recognition module to obtain the eighth data.

[0114] The pooling kernel size of the second maximum pooling layer is 2*2, and the stride parameter is 2.

[0115] S3204. The eighth data is normalized through the second batch normalization layer to obtain the target four-dimensional tensor data.

[0116] It is understandable that, since the intermediate four-dimensional tensor data has undergone one round of processing to obtain the target four-dimensional tensor data, the image width and height of the target four-dimensional tensor data is half the image width and height of the intermediate four-dimensional tensor data. Therefore, the data shape of the target four-dimensional tensor data is [A, 512, 1, 1], and the number of output images A is the same as the number of biological images in the original four-dimensional tensor data.

[0117] Step S330: The target four-dimensional tensor data is reshaped by the feature recognition module to obtain the same number of recognition features as the biological image.

[0118] Specifically, the target four-dimensional tensor data consists of several 1*1 pixel biological image data. Therefore, the target four-dimensional tensor data can be reshaped into two-dimensional recognition feature data. The format of the two-dimensional recognition feature data can be [E, F]. E can represent the number of images output by the feature recognition module, and F can represent the number of channels output by the feature recognition module, or the number of dimensional features of the biological image output by the feature recognition module.

[0119] The apparatus for recognizing biometric features provided in the embodiments of this application will be described below. The apparatus for recognizing biometric features described below can be referred to in correspondence with the method for recognizing biometric features described above.

[0120] See Figure 4 , Figure 4 This is a schematic diagram of a device for recognizing biometric features disclosed in an embodiment of this application.

[0121] like Figure 4 As shown, the device may include:

[0122] Tensor data acquisition unit 11 is used to acquire four-dimensional tensor data based on biological images. The initial four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height.

[0123] The feature recognition unit 12 is used to input the four-dimensional tensor data into a pre-established biometric recognition model and output recognition features with the same number as the number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

[0124] Optionally, the main structure of the model includes downsampling convolutional blocks;

[0125] The feature recognition unit includes:

[0126] The main structure recognition unit is used to perform five rounds of downsampling processing on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain four-dimensional tensor data to be recognized with the same number of images as the number of biological images. The number of channels in the four-dimensional tensor data to be recognized is greater than the number of input channels.

[0127] The identification module identification unit is used to perform feature identification on the four-dimensional tensor data to be identified through the feature identification module of the biometric identification model, and obtain identification features with the same number as the number of biological images, wherein the number of dimensions of each identification feature is greater than the number of channels of the four-dimensional tensor data to be identified.

[0128] Optionally, the image width and image height of the four-dimensional tensor data are both 256 pixels, and the number of input channels of the four-dimensional tensor data is 3;

[0129] The main structure identification unit includes:

[0130] The first main structure recognition subunit is used to perform the first round of downsampling processing on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain the first four-dimensional tensor data. The image width and image height of the first four-dimensional tensor data are both 128 pixels.

[0131] The second main structure recognition subunit is used to perform a second round of downsampling processing on the first four-dimensional tensor data through the downsampling convolution block to obtain the second four-dimensional tensor data. The image width and image height of the second four-dimensional tensor data are both 64 pixels.

[0132] The third main structure recognition subunit is used to perform a third round of downsampling processing on the second four-dimensional tensor data through the downsampling convolution block to obtain the third four-dimensional tensor data. The image width and image height of the third four-dimensional tensor data are both 32 pixels.

[0133] The fourth main structure recognition subunit is used to perform a fourth round of downsampling processing on the third four-dimensional tensor data through the downsampling convolution block to obtain the fourth four-dimensional tensor data, wherein the image width and image height of the fourth four-dimensional tensor data are both 16 pixels.

[0134] The fifth main structure recognition subunit is used to perform a fifth round of downsampling processing on the fourth four-dimensional tensor data through the downsampling convolution block to obtain the four-dimensional tensor data to be recognized. The image width and image height of the four-dimensional tensor data to be recognized are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be recognized is 384.

[0135] Optionally, the image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384.

[0136] The identification module identification unit includes:

[0137] The first recognition module recognition subunit is used to process the four-dimensional tensor data to be recognized in two rounds through the feature recognition module of the biometric recognition model according to the processing order of first convolution, then normalization, and finally maximum pooling, to obtain intermediate four-dimensional tensor data. The image width and image height of the intermediate four-dimensional tensor data are both 2 pixels.

[0138] The second recognition module recognition subunit is used to perform channel transformation convolution, normalization processing and max pooling processing on the intermediate four-dimensional tensor data through the feature recognition module to obtain target four-dimensional tensor data. The image width and image height of the target four-dimensional tensor data are both single pixels, and the number of dimensions of the target four-dimensional tensor data is 512.

[0139] The third identification module identification subunit is used to reshape the target four-dimensional tensor data through the feature identification module to obtain the same number of identification features as the biological image.

[0140] The biometric identification device provided in this application embodiment can be applied to biometric identification devices, such as terminals: mobile phones, computers, etc. Optionally, Figure 5 The diagram shows the hardware structure of a biometric identification device. (Refer to...) Figure 5 The hardware structure of a biometric identification device may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4;

[0141] In this embodiment, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4.

[0142] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0143] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0144] The memory stores a program, which the processor can call. The program is used for:

[0145] Obtain four-dimensional tensor data based on biological images. The initial four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height.

[0146] The four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features with the same number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

[0147] Optionally, the refined and extended functions of the program can be found in the description above.

[0148] This application embodiment also provides a storage medium that can store a program suitable for execution by a processor, the program being used for:

[0149] Obtain four-dimensional tensor data based on biological images. The initial four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height.

[0150] The four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features with the same number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

[0151] Optionally, the refined and extended functions of the program can be found in the description above.

[0152] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0153] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0154] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying biometric features, characterized in that, include: Obtain four-dimensional tensor data based on biological images, wherein the four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height; The four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features with the same number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer. The main structure of the model includes downsampling convolutional blocks; The four-dimensional tensor data is input into a pre-established biometric recognition model, which outputs recognition features equal in number to the number of biological images, including: By using the downsampling convolution block in the main structure of the pre-established biometric recognition model, the four-dimensional tensor data is downsampled five times to obtain four-dimensional tensor data to be recognized with the same number of images as the number of biological images. The number of channels in the four-dimensional tensor data to be recognized is greater than the number of input channels. The feature recognition module of the biometric recognition model performs feature recognition on the four-dimensional tensor data to be recognized, and obtains the same number of recognition features as the biological image. The number of dimensions of each recognition feature is greater than the number of channels of the four-dimensional tensor data to be recognized. The image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384. The feature recognition module of the biometric recognition model performs feature recognition on the four-dimensional tensor data to be identified, obtaining the same number of recognition features as the biological image, including: Following the processing order of convolution, normalization, and max pooling, the feature recognition module of the biometric recognition model processes the four-dimensional tensor data to be identified in two rounds to obtain intermediate four-dimensional tensor data. The image width and image height of the intermediate four-dimensional tensor data are both 2 pixels. The feature recognition module performs channel transformation convolution, normalization and max pooling on the intermediate four-dimensional tensor data to obtain the target four-dimensional tensor data. The image width and image height of the target four-dimensional tensor data are both single pixels, and the number of dimensions of the target four-dimensional tensor data is 512. The feature recognition module reshapes the target four-dimensional tensor data to obtain the same number of recognition features as the biological image.

2. The method according to claim 1, characterized in that, The image width and image height of the four-dimensional tensor data are both 256 pixels, and the number of input channels of the four-dimensional tensor data is 3; The four-dimensional tensor data is downsampled five times using downsampling convolutional blocks in the main structure of a pre-established biometric recognition model to obtain the four-dimensional tensor data to be recognized, including: The four-dimensional tensor data is downsampled for the first time by using the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain the first four-dimensional tensor data. The image width and image height of the first four-dimensional tensor data are both 128 pixels. The first four-dimensional tensor data is downsampled in the second round by the downsampling convolution block to obtain the second four-dimensional tensor data. The image width and image height of the second four-dimensional tensor data are both 64 pixels. The third round of downsampling processing is performed on the second four-dimensional tensor data by the downsampling convolution block to obtain the third four-dimensional tensor data. The image width and image height of the third four-dimensional tensor data are both 32 pixels. The fourth round of downsampling processing is performed on the third four-dimensional tensor data by the downsampling convolution block to obtain the fourth four-dimensional tensor data. The image width and image height of the fourth four-dimensional tensor data are both 16 pixels. The fourth four-dimensional tensor data is downsampled for the fifth time by the downsampling convolution block to obtain the four-dimensional tensor data to be identified. The image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384.

3. A biometric identification device, characterized in that, The device, used for biometric identification as described in claim 1, comprises: Tensor data acquisition unit is used to acquire four-dimensional tensor data based on biological images. The four-dimensional tensor data includes information on the number of biological images, the number of input channels, the image width, and the image height. The feature recognition unit is used to input the four-dimensional tensor data into a pre-established biometric recognition model and output the same number of recognition features as the number of biological images. The number of dimensions of each recognition feature is the number of output channels of the biometric recognition model, wherein the number of output channels is greater than the number of input channels. The model structure of the biometric recognition model consists of a main model structure and a feature recognition module. The main model structure is a MobileNetV2 structure, and the activation function used in the main model structure is the FReLU activation function. The feature recognition module is connected to the end of the main model structure. The feature recognition module includes a first convolutional layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second batch normalization layer, and a second max pooling layer.

4. The apparatus according to claim 3, characterized in that, The main structure of the model includes downsampling convolutional blocks; The feature recognition unit includes: The main structure recognition unit is used to perform five rounds of downsampling processing on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain four-dimensional tensor data to be recognized with the same number of images as the number of biological images. The number of channels in the four-dimensional tensor data to be recognized is greater than the number of input channels. The identification module identification unit is used to perform feature identification on the four-dimensional tensor data to be identified through the feature identification module of the biometric identification model, and obtain identification features with the same number as the number of biological images, wherein the number of dimensions of each identification feature is greater than the number of channels of the four-dimensional tensor data to be identified.

5. The apparatus according to claim 4, characterized in that, The image width and image height of the four-dimensional tensor data are both 256 pixels, and the number of input channels of the four-dimensional tensor data is 3; The main structure identification unit includes: The first main structure recognition subunit is used to perform the first round of downsampling processing on the four-dimensional tensor data through the downsampling convolution block in the main structure of the pre-established biometric recognition model to obtain the first four-dimensional tensor data. The image width and image height of the first four-dimensional tensor data are both 128 pixels. The second main structure recognition subunit is used to perform a second round of downsampling processing on the first four-dimensional tensor data through the downsampling convolution block to obtain the second four-dimensional tensor data. The image width and image height of the second four-dimensional tensor data are both 64 pixels. The third main structure recognition subunit is used to perform a third round of downsampling processing on the second four-dimensional tensor data through the downsampling convolution block to obtain the third four-dimensional tensor data. The image width and image height of the third four-dimensional tensor data are both 32 pixels. The fourth main structure recognition subunit is used to perform a fourth round of downsampling processing on the third four-dimensional tensor data through the downsampling convolution block to obtain the fourth four-dimensional tensor data, wherein the image width and image height of the fourth four-dimensional tensor data are both 16 pixels. The fifth main structure recognition subunit is used to perform a fifth round of downsampling processing on the fourth four-dimensional tensor data through the downsampling convolution block to obtain the four-dimensional tensor data to be recognized. The image width and image height of the four-dimensional tensor data to be recognized are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be recognized is 384.

6. The apparatus according to claim 4, characterized in that, The image width and image height of the four-dimensional tensor data to be identified are both 8 pixels, and the number of input channels of the four-dimensional tensor data to be identified is 384. The identification module identification unit includes: The first recognition module recognition subunit is used to process the four-dimensional tensor data to be recognized in two rounds through the feature recognition module of the biometric recognition model according to the processing order of first convolution, then normalization, and finally maximum pooling, to obtain intermediate four-dimensional tensor data. The image width and image height of the intermediate four-dimensional tensor data are both 2 pixels. The second recognition module recognition subunit is used to perform channel transformation convolution, normalization processing and max pooling processing on the intermediate four-dimensional tensor data through the feature recognition module to obtain target four-dimensional tensor data. The image width and image height of the target four-dimensional tensor data are both single pixels, and the number of dimensions of the target four-dimensional tensor data is 512. The third identification module identification subunit is used to reshape the target four-dimensional tensor data through the feature identification module to obtain the same number of identification features as the biological image.

7. A biometric identification device, characterized in that, Including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement the steps of the biometric identification method as described in any one of claims 1 and 2.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the biometric identification method as described in any one of claims 1 and 2.

Images