Small-area fingerprint identification method and system and storage medium

Abstract

The invention provides a small-area fingerprint identification method and system and a storage medium, and the method comprises the steps: obtaining a fingerprint image, and inputting the fingerprint image into a pre-trained fingerprint feature extraction network model; a block-by-block reasoning strategy is adopted for the fingerprint image through the fingerprint feature extraction network model to obtain a fingerprint low-dimensional feature map; obtaining a fingerprint high-dimensional feature map based on the low-dimensional feature map through a layer-by-layer reasoning strategy; processing the fingerprint high-dimensional feature map through a preset two-dimensional convolution layer, an average pooling layer and a full connection layer in sequence to obtain a final fingerprint feature vector; and calculating the similarity between the fingerprint image to be identified and the target fingerprint image through a cosine similarity function based on the final fingerprint feature vector, and completing fingerprint identification when a set threshold value is reached. Wherein the fingerprint feature extraction network model is obtained by training a fingerprint feature extraction neural network through a pre-constructed fingerprint identification data set. The problems that existing small-area fingerprint recognition is low in precision and low in speed are solved.

Description

Technical Field

[0001] The present invention relates to the field of fingerprint recognition technology, and in particular to a small-area fingerprint recognition method, system and storage medium. Background Art

[0002] Fingerprint recognition technology has been widely used in mobile devices, access control systems, payment verification and other fields. However, with the development trend of miniaturization of equipment, small-area fingerprint sensors have gradually become the mainstream choice. Due to the limitation of physical size, the fingerprint image area collected by such sensors is small and the feature information is limited. The fingerprint recognition method based on traditional minutiae and feature points has been relatively mature in large-area fingerprint recognition technology, but on small-area fingerprints, due to the small number of minutiae or feature points extracted, the performance of such algorithms on small-area fingerprints is greatly reduced. At the same time, the fingerprint acquisition image is affected by the sensor, the pressure of the finger, the dryness and wetness of the finger, the degree of contamination of the finger, etc. The traditional minutiae or feature point method has unstable extracted minutiae or feature points under various noises and interferences, and the robustness of the algorithm needs to be improved. In devices with limited computing resources such as embedded systems, complex feature extraction algorithms are difficult to run in real time, which seriously affects the user experience. In recent years, with the remarkable progress made by deep learning technology in the field of computer vision, its powerful feature learning ability has provided new ideas for solving the problem of small-area fingerprint recognition. Summary of the invention

[0003] The present invention provides a small-area fingerprint recognition method, system and storage medium, which are used to solve the problems of low precision and slow speed of existing small-area fingerprint recognition.

[0004] The present invention provides a small-area fingerprint recognition method, comprising: Acquire a fingerprint image, and input the fingerprint image into a pre-trained fingerprint feature extraction network model; Using the fingerprint feature extraction network model to apply a block-by-block reasoning strategy to the fingerprint image, a fingerprint low-dimensional feature map is obtained; Based on the low-dimensional feature map, a fingerprint high-dimensional feature map is obtained through a layer-by-layer reasoning strategy; The fingerprint high-dimensional feature map is processed in sequence by a preset two-dimensional convolution layer, an average pooling layer and a fully connected layer to obtain a final fingerprint feature vector; Based on the final fingerprint feature vector, the similarity between the fingerprint image to be identified and the target fingerprint image is calculated by using a cosine similarity function, and the fingerprint is identified when a set threshold is reached; The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network with a pre-constructed fingerprint recognition data set.

[0005] According to a small-area fingerprint recognition method provided by the present invention, the fingerprint feature extraction network model is obtained by training a fingerprint feature extraction neural network through a pre-constructed fingerprint recognition data set, specifically comprising: Collect small-area fingerprint image data, calculate the positional relationship between every two images in the fingerprint image of the same finger, record and save the intersecting image pairs, and build a data set; Based on the data set, the fingerprint feature extraction neural network is iteratively trained in a supervised manner according to the defined feature distance measurement method, loss function and optimization algorithm to obtain a fingerprint feature extraction network model.

[0006] According to a small-area fingerprint recognition method provided by the present invention, during the process of deploying the fingerprint feature extraction network model to an embedded scenario, the fingerprint feature extraction network model is quantized and then deployed.

[0007] According to a small-area fingerprint recognition method provided by the present invention, before obtaining a fingerprint low-dimensional feature map by using a block-by-block reasoning strategy on the fingerprint image through the fingerprint feature extraction network model, the method includes: Inputting the fingerprint image into a fingerprint feature extraction network model, and capturing the global features and local features of the fingerprint through a fingerprint global and local feature combination unit in the fingerprint feature extraction network model; The global features and local features of the fingerprint are fused to generate fused features.

[0008] According to a small-area fingerprint recognition method provided by the present invention, the fingerprint feature extraction network model uses a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map on the fingerprint image, specifically comprising: Based on the fusion features, a fingerprint low-dimensional feature extraction unit in the fingerprint feature extraction network model is used to obtain a fingerprint low-dimensional feature map by adopting a block-by-block reasoning strategy; The low-dimensional feature extraction unit is generated by cascading multiple GB modules, each GB module includes multiple depth separation convolution layers, and the channels of the feature map are expanded and reduced through the multiple depth separation convolution layers.

[0009] According to a small-area fingerprint recognition method provided by the present invention, the fingerprint high-dimensional feature map is obtained by a layer-by-layer reasoning strategy based on the low-dimensional feature map, specifically comprising: Based on the low-dimensional feature graph, a fingerprint high-dimensional feature graph is obtained by using a fingerprint high-dimensional feature extraction unit in a fingerprint feature extraction network model, using a layer-by-layer reasoning strategy; The high-dimensional feature extraction unit is generated by cascading multiple MB modules, and each MB module includes a deep separation convolution, a point-by-point convolution and an inverted residual network structure.

[0010] The present invention also provides a small-area fingerprint recognition system, the system comprising: An image acquisition module is used to acquire a fingerprint image and input the fingerprint image into a pre-trained fingerprint feature extraction network model; A low-dimensional feature acquisition module, used for obtaining a fingerprint low-dimensional feature map by applying a block-by-block reasoning strategy to the fingerprint image through the fingerprint feature extraction network model; A high-dimensional feature acquisition module, used to obtain a fingerprint high-dimensional feature map based on the low-dimensional feature map through a layer-by-layer reasoning strategy; A feature vector acquisition module is used to process the fingerprint high-dimensional feature map through a preset two-dimensional convolution layer, an average pooling layer and a fully connected layer in sequence to obtain a final fingerprint feature vector; A recognition module, used to calculate the similarity between the fingerprint image to be recognized and the target fingerprint image based on the final fingerprint feature vector through a cosine similarity function, and complete the recognition of the fingerprint when a set threshold is reached; Among them, the fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network through a pre-constructed fingerprint recognition data set. The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the small-area fingerprint recognition methods described above.

[0011] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the small-area fingerprint recognition method described above is implemented.

[0012] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned small-area fingerprint recognition methods.

[0013] The present invention provides a small-area fingerprint recognition method, system and storage medium. Through a fingerprint feature extraction network model, a more stable and reliable feature vector is automatically extracted from a small-area fingerprint image, thereby overcoming the problem of insufficient features of traditional methods, thereby improving the accuracy of fingerprint matching and reducing the error rate and rejection rate; through the feature representation capability of the fingerprint feature extraction network model, the algorithm's resistance to adverse factors such as noise, blur, deformation, etc. is improved, ensuring that high-quality fingerprint features are obtained under various complex situations and enhancing the robustness of the algorithm; a lightweight deep learning model is designed so that it can achieve efficient fingerprint recognition even with limited computing resources, improve real-time performance and user experience, and thus make it suitable for a variety of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 This is one of the flow charts of a small-area fingerprint recognition method provided by the present invention.

[0016] Figure 2 It is a schematic diagram of the fingerprint feature extraction network model structure provided by the present invention.

[0017] Figure 3 It is a schematic diagram of the GB module structure provided by the present invention.

[0018] Figure 4 The present invention provides a schematic diagram of the module connection of a small-area fingerprint recognition system.

[0019] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention.

[0020] Reference numerals: 110: image acquisition module; 120: low-dimensional feature acquisition module; 130: high-dimensional feature acquisition module; 140: feature vector acquisition module; 150: recognition module; 510: processor; 520: communication interface; 530: memory; 540: communication bus. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Combine the following Figure 1 and Figure 2 , describing a small-area fingerprint recognition method of the present invention, characterized in that it includes: step 100, obtaining a fingerprint image, and inputting the fingerprint image into a pre-trained fingerprint feature extraction network model.

[0023] The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network with a pre-constructed fingerprint recognition data set.

[0024] In the present invention, a fingerprint feature extraction network is constructed, a fingerprint recognition data set is prepared, and a supervision strategy is used to train the constructed fingerprint feature extraction network model; a target model of the trained fingerprint feature extraction network is deployed, and when the deployment environment is a resource-constrained embedded scenario, the network model is first quantized, and then the quantized target model is deployed; model inference is performed to extract fingerprint features, and in the model inference stage, the model inference strategy is optimized to reduce the memory occupancy of the model in the inference stage; and the target fingerprint is recognized through distance measurement and similarity calculation.

[0025] The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network through the pre-built fingerprint recognition data set, which specifically includes: Collect small-area fingerprint image data, calculate the positional relationship between every two images in the fingerprint image of the same finger, record and save the intersecting image pairs, and build a data set; Based on the data set, the fingerprint feature extraction neural network is iteratively trained in a supervised manner according to the defined feature distance measurement method, loss function and optimization algorithm to obtain a fingerprint feature extraction network model.

[0026] Specifically, we construct a data set, collect small-area fingerprint image data, calculate the positional relationship between every two images of the same finger's fingerprint image, and record and save the intersecting image pairs for subsequent reading and training of the network model. The supervised training of the model uses a triplet method to read, and each group of image pairs uses 3 images.<a,p,n> is used for network training, where a represents the selected anchor point, p represents the sample with the same finger as a and the area intersection, and n represents the sample with a different finger from a.

[0027] Define the feature distance measurement method, loss function and optimization algorithm in the feature extraction network model training process, and then implement iterative training of the feature extraction network through supervision. Deploy the trained fingerprint feature extraction network model. If the deployment scenario is embedded, quantize the model and deploy the quantized target model.

[0028] Step 200: Using the fingerprint feature extraction network model to perform a block-by-block reasoning strategy on the fingerprint image to obtain a fingerprint low-dimensional feature map.

[0029] Specifically, it includes: based on the fusion feature, through the fingerprint low-dimensional feature extraction unit in the fingerprint feature extraction network model, adopting a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map; The low-dimensional feature extraction unit is generated by cascading multiple GB modules, each GB module includes multiple depth separation convolution layers, and the channels of the feature map are expanded and reduced through the multiple depth separation convolution layers.

[0030] In the present invention, a fingerprint feature extraction network model is provided with a fingerprint global and local feature combination unit, a low-dimensional feature extraction unit, a high-dimensional feature extraction unit and a fingerprint feature output unit.

[0031] The fingerprint global and local feature combination unit in the fingerprint feature extraction network model includes feature map segmentation, sliding window block extraction, feature map scaling interpolation, feature map splicing module (UFM) and channel attention module (CAM). In the fingerprint matching stage in the fingerprint application scenario, the paired fingerprint maps will overlap to varying degrees, and in small-area fingerprints, the overlap is generally small, which requires that the extracted fingerprint features can not only contain overall features, but also local features. The block, sliding window, and scaling combination operations can enable the subsequent network to better capture the global and local features of the fingerprint. The channel attention module CAM is used to capture the weights of each block and global feature in the synthesized fingerprint feature.

[0032] In the present invention, the number of blocks of the UFM module should be moderate. Too few blocks are not easy to extract local features better, and too many blocks will increase the amount of calculation. Here, it is necessary to adjust according to the scale of the input original fingerprint image, generally between 16-32 blocks. At the same time, in the block process, it is necessary to ensure that the width and height of the block can be divided by the width and height of the feature map. In the process of sliding window block extraction by the UFM module, the obtained area should overlap, and an overlapping sliding window block extraction strategy is adopted. Before the block features are concat, the global feature map and each block feature map need to be normalized to the same scale. Here, bilinear interpolation is used for normalization.

[0033] refer to Figure 3The low-dimensional feature extraction unit is composed of multiple GB modules cascaded together, which is used to extract low-dimensional fingerprint features such as low-level texture of fingerprints. The GB module includes 4 deep separation convolutions (DWConv), which reduce the model size and computational complexity while ensuring a certain accuracy. The GB model adopts an inverted residual network structure. The first two DWConvs expand the channels of the feature map, and the last two DWConvs reduce the channels of the feature map. The low-dimensional features of fingerprints are mainly related basic features such as fingerprint texture and lines. Among these low-dimensional features of fingerprints, there are many similar feature maps and a large amount of redundant information. In the fingerprint field, these similar redundant information is more conducive to the subsequent network to extract the final features. When designing the GB module, the construction of redundant information is fully considered. First, the feature map is passed through the first DWConv to obtain a new feature map. The new feature map is input into the second DWConv to realize the construction of redundant information, and then the two feature maps are combined into a new feature map by Concat operation. The third and fourth DWconv are similar to the first and second. Among the four DWConv operations of the GB module, except for the first DWConv operation whose stride value is 1 or 2, the strides of the other three DWConv operations are all 1.

[0034] In the present invention, the low-dimensional feature extraction unit is composed of multiple cascaded GB modules. In order to reduce the amount of calculation and model size, and reduce the model's occupancy of memory and other resources, the number of GB modules should not be too large, generally 3-5. The DWConv convolution kernel size in the GB module is 3x3, and it adopts an inverted residual network structure. The expansion ratio of the feature map channel of the GB module is selected to be 3-6.

[0035] Step 300: Obtain a fingerprint high-dimensional feature map based on the low-dimensional feature map through a layer-by-layer reasoning strategy.

[0036] Specifically, it includes: based on the low-dimensional feature map, a fingerprint high-dimensional feature extraction unit in a fingerprint feature extraction network model is used to obtain a fingerprint high-dimensional feature map by adopting a layer-by-layer reasoning strategy; The high-dimensional feature extraction unit is generated by cascading multiple MB modules, and each MB module includes a deep separation convolution, a point-by-point convolution and an inverted residual network structure.

[0037] The high-dimensional feature extraction unit is composed of multiple MB modules in cascade, and the MB module is mainly composed of deep separation convolution, point-by-point convolution, and inverted residual network structure. It has low computational complexity while ensuring model accuracy.

[0038] In the present invention, the high-dimensional feature extraction unit is formed by cascading multiple MB modules. The DWConv in the MB module uses a 3x3 convolution kernel with a stride of 1 or 2. When the stride is 1, the scale of the input and output feature maps of the MB module remains unchanged. When the stride is 2, the MB module reduces the scale of the input feature map. The number of MB modules with a stride of 2 is generally 1-2. Too few MB modules with a stride of 1 will result in low model accuracy. A large number of MB modules will increase the model size and the amount of calculation. Here, the number of MB modules is adjusted between 5 and 12.

[0039] The extracted high-dimensional fingerprint features are sequentially subjected to Conv2d 1x1, AvgPool, and FC operations to obtain the final fingerprint features.

[0040] Step 400: Process the fingerprint high-dimensional feature map in sequence through a preset two-dimensional convolution layer, an average pooling layer, and a fully connected layer to obtain a final fingerprint feature vector.

[0041] In the present invention, the fingerprint feature output part is composed of a 1x1 Conv2d (two-dimensional convolutional layer), an AvgPool (average pooling layer) and a FC layer (fully connected layer) cascaded in sequence, and the final output fingerprint feature length is set by the FC layer.

[0042] The final feature extraction network model is determined according to different fingerprint image scales and corresponding hardware resources. It can be manually adjusted using experience, or a MnasNet-like search space method can be used to search for the optimal model structure.

[0043] In the present invention, in the reasoning stage of the fingerprint feature extraction network model, the common reasoning method is to run in a layer-by-layer manner, that is, each operation such as convolution allocates memory space according to the scale of the feature map, and then completes the relevant calculation operations of the feature map layer by layer. In the front part of the feature extraction network, the model before the fingerprint high-dimensional feature extraction unit, due to the large scale of the fingerprint feature map, the peak memory is large when the traditional reasoning method is used, which is not conducive to deployment in memory-constrained scenarios such as embedded systems. In order to alleviate this problem, a block-by-block reasoning strategy is designed, that is, the relevant calculations of the feature map are completed in a block-by-block manner in the low-dimensional feature extraction unit and the previous part, and the remaining part is still performed in a layer-by-layer manner. Since the reasoning is performed in a block-by-block manner, the larger the receptive field will lead to more overlapping areas and repeated calculations. Therefore, in the design of the algorithm model, a smaller receptive field convolution kernel and fewer convolution modules are used in the initial part to reduce the overlapping area and the amount of calculation, and the convolution modules are appropriately increased in the later part to compensate for the loss of accuracy.

[0044] Step 500: Based on the final fingerprint feature vector, the similarity between the fingerprint image to be identified and the target fingerprint image is calculated by using a cosine similarity function, and the fingerprint is identified when a set threshold is reached.

[0045] Specifically, in the present invention, the process of fingerprint recognition by using a fingerprint feature extraction network model includes: Construct a data set, collect small-area fingerprint image data, calculate the positional relationship between every two images in the fingerprint image of the same finger, and record and save the intersecting image pairs for subsequent reading and training of the network model. The supervised training of the model uses the triplet method to read, and each group of image pairs uses 3<a,p,n> is used for network training, where a represents the selected anchor point, p represents the sample with the same finger as a and the area intersection, and n represents the sample with a different finger from a; Define the feature distance measurement method, loss function and optimization algorithm in the feature extraction network model training process, and then implement iterative training of the feature extraction network through supervision; Deploy the trained fingerprint feature extraction network model. If the deployment scenario is an embedded scenario, quantize the model and deploy the quantized target model. Optimize the model reasoning strategy. In the first Conv2d of the algorithm model, the fingerprint global and local feature combination unit, and the fingerprint low-dimensional feature extraction unit, a block-by-block reasoning strategy is adopted to calculate and obtain the low-dimensional feature map of the fingerprint piece by piece, thereby reducing the memory usage. Define the similarity calculation method between fingerprint features. Here, cosine similarity is used to calculate the similarity between two fingerprints, and then the threshold method is used to identify whether the two fingerprints are from the same finger, thereby achieving the purpose of fingerprint recognition.

[0046] In a specific embodiment, when the fingerprint feature extraction network model faces different hardware resources and fingerprint images of different scales, the final model can be manually designed based on human experience, or the final model can be determined using a MnasNet-like search space method.

[0047] For example, under 512KB SRAM and 1MB Flash MCU, the input fingerprint image scale is 80x64, and the designed feature extraction network model structure is shown in Table 1. Input is the input feature scale, Operator is the operation module, expand is the channel expansion ratio inside the module, Output is the output channel scale, and stride is the relevant step size setting inside the module.

[0048] Table 1 First, construct a data set and use a fingerprint sensor to collect image data of small-area fingerprints. Save the fingerprint images of the same finger in the same folder, and save the fingerprint images of different fingers in different folders.

[0049] Data preprocessing is performed. The Fourier-Mellin transform algorithm is used to cyclically calculate the positional relationship between every two images of all fingerprint images of the same finger. The names of fingerprint image pairs whose cross-area is greater than 1 / 4 of the fingerprint image are saved. The names of the image pairs are saved in text for the subsequent construction of triplet data for model training. In the data reading for model iterative training, 3 images are used for each group of image pairs.<a,p,n> Fingerprint image, where a represents the selected anchor point, p represents the sample with the same finger as a and the area of ​​intersection, and n represents the sample with a different finger from a. Each time the data is read, a finger is randomly selected and a pair of images of the finger saved above is randomly selected to form<a,p> Then randomly extract a fingerprint from other fingers to form the final<a,p,n> Fingerprint images are relatively. In order to enhance data diversity, during the model training data reading phase, the read-in fingerprint image data is randomly enhanced, and the brightness, contrast, image quality, random noise, and rotation transformation online enhancement strategies are randomly performed. The angle of the random rotation strategy should not be too large, otherwise it will cause a large change in the overlapping area.

[0050] Construct the required fingerprint feature extraction network model, and construct three binary classification heads to form a training network. The three classification heads are used to judge<a,p> ,<a,n> ,<p,n> Whether they belong to the same category, that is, whether they belong to the same finger.

[0051] Define the feature distance measurement method, loss function and optimization algorithm in the feature extraction network model training process, and implement the training of the feature extraction network model through supervised training.

[0052] The loss function is defined as follows: in, is the cosine similarity, is the triplet loss, is the cross entropy loss, is the final loss function, , 2. 3 are<a,p> ,<a,n> ,<p,n> The cross entropy loss is , , , Are the weight coefficients of the four losses, and here the four weight coefficients are selected as 0.3, 0.2, 0.2, and 0.2 respectively.

[0053] The AdamW optimizer was selected for iterative optimization, the initial learning rate was selected as 0.0001, the cosine annealing learning rate adjustment strategy was adopted, and the network model was trained iteratively step by step.

[0054] According to the actual deployment scenario, the trained feature extraction network model is quantized to further reduce the size of the model and the amount of calculation. Here, the parameters in the network model are quantized from 32 bits to 8 bits, and the quantized target model is deployed.

[0055] In the model inference stage, a block-by-block inference strategy is used to perform inference calculations on low-dimensional features, and a layer-by-layer inference strategy is used to perform inference calculations on high-dimensional features. Here, the block size is selected as 8x8, which can effectively reduce the memory usage in the inference stage.

[0056] The fingerprint image is inferred and calculated by the feature extraction network model to extract the feature vector of the fingerprint image. Then, the cosine similarity function is used to calculate the similarity between the fingerprint image to be identified and the target fingerprint image, and the fingerprint is recognized by the threshold method.

[0057] After testing, we found that the pain points faced by this field can be better solved, and the accuracy of small-area embedded fingerprints has been further improved, and the robustness of the algorithm has been enhanced. For example, in the evaluation of the test data set collected by our own multi-sensor, the FAR is less than one millionth at the 160x160 scale, and the FRR reaches 0.2%. At the 80x64 scale, the FAR is less than one millionth, and the FRR reaches an accuracy of less than 1%.

[0058] refer to Figure 4 The present invention also discloses a small-area fingerprint recognition system, the system comprising: An image acquisition module 110 is used to acquire a fingerprint image and input the fingerprint image into a pre-trained fingerprint feature extraction network model; A low-dimensional feature acquisition module 120, configured to obtain a fingerprint low-dimensional feature map by applying a block-by-block reasoning strategy to the fingerprint image through the fingerprint feature extraction network model; A high-dimensional feature acquisition module 130, configured to obtain a fingerprint high-dimensional feature map based on the low-dimensional feature map by a layer-by-layer reasoning strategy; The feature vector acquisition module 140 is used to process the fingerprint high-dimensional feature map through a preset two-dimensional convolution layer, an average pooling layer and a fully connected layer in sequence to obtain a final fingerprint feature vector; The recognition module 150 is used to calculate the similarity between the fingerprint image to be recognized and the target fingerprint image based on the final fingerprint feature vector through the cosine similarity function, and complete the recognition of the fingerprint when a set threshold is reached; The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network with a pre-constructed fingerprint recognition data set.

[0059] The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network through a pre-constructed fingerprint recognition data set, specifically including: Collect small-area fingerprint image data, calculate the positional relationship between every two images in the fingerprint image of the same finger, record and save the intersecting image pairs, and build a data set; Based on the data set, the fingerprint feature extraction neural network is iteratively trained in a supervised manner according to the defined feature distance measurement method, loss function and optimization algorithm to obtain a fingerprint feature extraction network model.

[0060] When the fingerprint feature extraction network model is deployed to an embedded scenario, the fingerprint feature extraction network model is quantized and then deployed.

[0061] Before the fingerprint feature extraction network model is used to extract the fingerprint image using a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map, the method includes: Inputting the fingerprint image into a fingerprint feature extraction network model, and capturing the global features and local features of the fingerprint through a fingerprint global and local feature combination unit in the fingerprint feature extraction network model; The global features and local features of the fingerprint are fused to generate fused features.

[0062] The method of using the fingerprint feature extraction network model to extract the fingerprint image using a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map specifically includes: Based on the fusion features, a fingerprint low-dimensional feature extraction unit in the fingerprint feature extraction network model is used to obtain a fingerprint low-dimensional feature map by adopting a block-by-block reasoning strategy; The low-dimensional feature extraction unit is generated by cascading multiple GB modules, each GB module includes multiple depth separation convolution layers, and the channels of the feature map are expanded and reduced through the multiple depth separation convolution layers.

[0063] The step of obtaining a fingerprint high-dimensional feature map based on the low-dimensional feature map by a layer-by-layer reasoning strategy specifically includes: Based on the low-dimensional feature graph, a fingerprint high-dimensional feature graph is obtained by using a fingerprint high-dimensional feature extraction unit in a fingerprint feature extraction network model, using a layer-by-layer reasoning strategy; The high-dimensional feature extraction unit is generated by cascading multiple MB modules, and each MB module includes a deep separation convolution, a point-by-point convolution and an inverted residual network structure.

[0064] A small-area fingerprint recognition system provided by the present invention automatically extracts more stable and reliable feature vectors from a small-area fingerprint image through a fingerprint feature extraction network model, thereby overcoming the problem of insufficient features in traditional methods, thereby improving the accuracy of fingerprint matching and reducing the error rate and rejection rate; through the feature representation capability of the fingerprint feature extraction network model, the algorithm's resistance to adverse factors such as noise, blur, and deformation is improved, ensuring that high-quality fingerprint features are obtained under various complex situations and enhancing the robustness of the algorithm; a lightweight deep learning model is designed to enable efficient fingerprint recognition to be achieved even with limited computing resources, thereby improving real-time performance and user experience, thereby making it suitable for a variety of application scenarios.

[0065] Figure 5 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 5 As shown, the electronic device may include: a processor (processor) 510, a communication interface (Communications Interface) 520, a memory (memory) 530 and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call the logic instructions in the memory 530 to execute a small-area fingerprint recognition method, which includes: obtaining a fingerprint image, inputting the fingerprint image into a pre-trained fingerprint feature extraction network model; using the fingerprint feature extraction network model to use a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map for the fingerprint image; based on the low-dimensional feature map, a fingerprint high-dimensional feature map is obtained by using a layer-by-layer reasoning strategy; the fingerprint high-dimensional feature map is processed in sequence by a preset two-dimensional convolution layer, an average pooling layer, and a fully connected layer to obtain a final fingerprint feature vector; based on the final fingerprint feature vector, the similarity between the fingerprint image to be identified and the target fingerprint image is calculated by using a cosine similarity function to complete the identification of the fingerprint when a set threshold is reached; wherein the fingerprint feature extraction network model is obtained by training a fingerprint feature extraction neural network using a pre-constructed fingerprint recognition data set.

[0066] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0067] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a small-area fingerprint recognition method provided by the above methods, the method including: obtaining a fingerprint image, inputting the fingerprint image into a pre-trained fingerprint feature extraction network model; using the fingerprint feature extraction network model to obtain a fingerprint low-dimensional feature map using a block-by-block reasoning strategy for the fingerprint image; obtaining a fingerprint high-dimensional feature map based on the low-dimensional feature map using a layer-by-layer reasoning strategy; processing the fingerprint high-dimensional feature map through a preset two-dimensional convolutional layer, an average pooling layer and a fully connected layer in sequence to obtain a final fingerprint feature vector; based on the final fingerprint feature vector, calculating the similarity between the fingerprint image to be identified and the target fingerprint image through a cosine similarity function, and completing the identification of the fingerprint when the set threshold is reached; wherein the fingerprint feature extraction network model is obtained by training a fingerprint feature extraction neural network using a pre-constructed fingerprint recognition data set.

[0068] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute a small-area fingerprint recognition method provided by the above methods, the method comprising: obtaining a fingerprint image, inputting the fingerprint image into a pre-trained fingerprint feature extraction network model; obtaining a fingerprint low-dimensional feature map by adopting a block-by-block reasoning strategy for the fingerprint image through the fingerprint feature extraction network model; obtaining a fingerprint high-dimensional feature map based on the low-dimensional feature map through a layer-by-layer reasoning strategy; processing the fingerprint high-dimensional feature map through a preset two-dimensional convolutional layer, an average pooling layer and a fully connected layer in sequence to obtain a final fingerprint feature vector; calculating the similarity between the fingerprint image to be identified and the target fingerprint image through a cosine similarity function based on the final fingerprint feature vector, and completing the recognition of the fingerprint by reaching a set threshold; wherein the fingerprint feature extraction network model is obtained by training a fingerprint feature extraction neural network through a pre-constructed fingerprint recognition data set.

[0069] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0070] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A small area fingerprint recognition method, characterized in that: include: Acquire a fingerprint image, and input the fingerprint image into a pre-trained fingerprint feature extraction network model; Using the fingerprint feature extraction network model to apply a block-by-block reasoning strategy to the fingerprint image, a fingerprint low-dimensional feature map is obtained; Based on the low-dimensional feature map, a fingerprint high-dimensional feature map is obtained through a layer-by-layer reasoning strategy; The fingerprint high-dimensional feature map is processed in sequence by a preset two-dimensional convolution layer, an average pooling layer and a fully connected layer to obtain a final fingerprint feature vector; Based on the final fingerprint feature vector, the similarity between the fingerprint image to be identified and the target fingerprint image is calculated by using a cosine similarity function, and the fingerprint is identified when a set threshold is reached; The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network with a pre-constructed fingerprint recognition data set.

2. The small-area fingerprint recognition method according to claim 1, characterized in that: The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network through a pre-constructed fingerprint recognition data set, specifically including: Collect small-area fingerprint image data, calculate the positional relationship between every two images in the fingerprint image of the same finger, record and save the intersecting image pairs, and build a data set; Based on the data set, the fingerprint feature extraction neural network is iteratively trained in a supervised manner according to the defined feature distance measurement method, loss function and optimization algorithm to obtain a fingerprint feature extraction network model.

3. The small-area fingerprint recognition method according to claim 2, characterized in that: When the fingerprint feature extraction network model is deployed to an embedded scenario, the fingerprint feature extraction network model is quantized and then deployed.

4. The small-area fingerprint recognition method according to claim 1, characterized in that: Before the fingerprint feature extraction network model is used to extract the fingerprint image using a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map, the method includes: Inputting the fingerprint image into a fingerprint feature extraction network model, and capturing the global features and local features of the fingerprint through a fingerprint global and local feature combination unit in the fingerprint feature extraction network model; The global features and local features of the fingerprint are fused to generate fused features.

5. The small-area fingerprint recognition method according to claim 1, characterized in that: The method of using the fingerprint feature extraction network model to extract the fingerprint image using a block-by-block reasoning strategy to obtain a fingerprint low-dimensional feature map specifically includes: Based on the fusion features, a fingerprint low-dimensional feature extraction unit in the fingerprint feature extraction network model is used to obtain a fingerprint low-dimensional feature map by adopting a block-by-block reasoning strategy; The low-dimensional feature extraction unit is generated by cascading multiple GB modules, each GB module includes multiple depth separation convolution layers, and the channels of the feature map are expanded and reduced through the multiple depth separation convolution layers.

6. The small-area fingerprint recognition method according to claim 1, characterized in that: The step of obtaining a fingerprint high-dimensional feature map based on the low-dimensional feature map by a layer-by-layer reasoning strategy specifically includes: Based on the low-dimensional feature graph, a fingerprint high-dimensional feature extraction unit in a fingerprint feature extraction network model is used to obtain a fingerprint high-dimensional feature graph by adopting a layer-by-layer reasoning strategy; The high-dimensional feature extraction unit is generated by cascading multiple MB modules, and each MB module includes a deep separation convolution, a point-by-point convolution and an inverted residual network structure.

7. A small area fingerprint recognition system, characterized in that: The system comprises: An image acquisition module is used to acquire a fingerprint image and input the fingerprint image into a pre-trained fingerprint feature extraction network model; A low-dimensional feature acquisition module, used for obtaining a fingerprint low-dimensional feature map by applying a block-by-block reasoning strategy to the fingerprint image through the fingerprint feature extraction network model; A high-dimensional feature acquisition module, used to obtain a fingerprint high-dimensional feature map based on the low-dimensional feature map through a layer-by-layer reasoning strategy; A feature vector acquisition module is used to process the fingerprint high-dimensional feature map through a preset two-dimensional convolution layer, an average pooling layer and a fully connected layer in sequence to obtain a final fingerprint feature vector; A recognition module, used to calculate the similarity between the fingerprint image to be recognized and the target fingerprint image based on the final fingerprint feature vector through a cosine similarity function, and complete the recognition of the fingerprint when a set threshold is reached; The fingerprint feature extraction network model is obtained by training the fingerprint feature extraction neural network with a pre-constructed fingerprint recognition data set.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the small-area fingerprint recognition method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the small-area fingerprint recognition method according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the small-area fingerprint recognition method according to any one of claims 1 to 6 is implemented.

Images