Fingerprint identification method based on deep learning

Abstract

The invention discloses a fingerprint identification method based on deep learning, and the method comprises the following steps: S1, collecting and preprocessing a fingerprint image, and generating a standardized image; s2, constructing a fingerprint identification model which comprises a feature extraction module, a feature fusion module and a classification module; s3, using a whale optimization algorithm to jointly optimize a fingerprint identification model structure and training parameters; s4, configuring and training a model, and optimizing the weight by adopting a category boundary loss function; and S5, deploying a fingerprint identification model, receiving the real-time image, completing identification, and outputting a result and confidence. By fusing the residual network, the feature pyramid structure and the whale optimization algorithm, the precision and real-time performance of fingerprint identification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biometric recognition and image processing, and particularly relates to a fingerprint recognition method based on deep learning. BACKGROUND

[0002] Under the background of the increasing popularity of information security and identity authentication technology, fingerprint recognition as an important biometric recognition means is widely used in various scenarios such as financial payment, access control system, mobile terminal unlocking, etc. due to its uniqueness, convenience and stability. At present, the mainstream fingerprint recognition method is mostly based on the deep convolutional neural network (CNN) architecture, which extracts local ridge features in the fingerprint image for identity discrimination. However, the traditional convolutional structure often has the problems of incomplete feature extraction and insufficient feature expression capability when processing fingerprint images of different quality and different scales, especially in blurred, partially occluded or low-contrast fingerprint images, the recognition effect is significantly reduced. In addition, the fixed model structure and the manually set parameters also limit the adaptability and expansibility of the model.

[0003] In order to enhance the discrimination ability of the model for complex fingerprint images, existing research has introduced a multi-scale feature fusion strategy, such as the feature pyramid network (FPN), to extract ridge details and structural information at different resolutions and improve the robustness of the model. However, the existing multi-scale fusion method generally uses a fixed fusion rule, lacks screening of redundant features and adaptive strengthening of effective features, and is prone to information redundancy or feature imbalance. At the same time, in the design of deep network structure, uniform channel configuration and symmetric residual structure are still commonly used, which cannot be personalized adjusted for different levels of features, resulting in increased redundant calculation of the model and lack of flexibility.

[0004] In addition, the performance of the fingerprint recognition model depends largely on the reasonable configuration of the structure parameters and the training parameters, and the existing methods mostly use manual experience to adjust the parameters, which is low in efficiency and poor in controllability, and it is difficult to balance the recognition accuracy and model compression. In view of the above problems, there is still a lack of a deep fingerprint recognition method that can integrate channel compression control, asymmetric residual connection and scale perception fusion mechanism, and automatically adjust the parameters combined with intelligent optimization algorithm.

[0005] Therefore, how to provide a fingerprint recognition method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to propose a fingerprint recognition method based on deep learning, which fully utilizes the feature extraction capability of deep residual neural network, the multi-scale fusion capability of feature pyramid network and the self-adaptive adjustment capability of intelligent optimization algorithm to model structure and parameters, and describes in detail the complete process of fingerprint image from collection, preprocessing, feature extraction, structure optimization to online recognition. The method introduces dynamic channel compression, asymmetric residual connection and scale perception gating mechanism, and configures the key structure parameters and training parameters through whale optimization algorithm, which has the advantages of high recognition accuracy, high model calculation efficiency and strong adaptability to complex image scenes, and is especially suitable for fingerprint recognition tasks in mobile terminals, embedded recognition systems and high security scenes.

[0007] According to the fingerprint recognition method based on deep learning, the method comprises the following steps:

[0008] S1, collecting fingerprint image data, preprocessing the fingerprint image data to obtain standardized fingerprint images;

[0009] S2, constructing a fingerprint recognition model, the fingerprint recognition model comprising a feature extraction module, a feature fusion module and a classification module, wherein the feature extraction module adopts a residual neural network-based backbone network to extract basic features of the fingerprint image, the feature fusion module introduces a feature pyramid network structure to upsample and horizontally connect the basic features to generate multi-scale fusion feature maps, and the classification module is used to input the multi-scale fusion features into a fully connected layer to output a fingerprint identity recognition result;

[0010] S3, jointly optimizing the structure parameters and training parameters of the fingerprint recognition model based on a whale optimization algorithm to obtain an optimal parameter combination;

[0011] S4, configuring the fingerprint recognition model with the optimal parameter combination, inputting the standardized fingerprint images for training, optimizing the model weights by using a class boundary-based metric learning loss function, and completing the fingerprint recognition model training process;

[0012] S5, deploying the trained fingerprint recognition model in a target system, receiving real-time collected fingerprint images, executing image input, feature extraction, feature fusion and identity recognition processes, outputting recognition results and confidence scores, and realizing online execution of the fingerprint recognition task.

[0013] Optionally, the fingerprint image data specifically comprises static gray-scale images and dynamic infrared images.

[0014] Optionally, the preprocessing of the fingerprint image data specifically comprises image gray-scale normalization, noise filtering, geometric alignment and ridge enhancement operations.

[0015] Optionally, S2 specifically comprises:

[0016] S21, the overall structure of the fingerprint identification model is constructed, including a feature extraction module, a feature fusion module and a classification module, the input is a standardized fingerprint image, and the output is a fingerprint identity recognition result;

[0017] S22, in the feature extraction module, a backbone structure based on a residual neural network is used to receive the standardized fingerprint image input, and a plurality of residual units are used to output a plurality of layers of basic feature maps C2, C3, C4 and C5 in turn, for representing fingerprint line features at different levels;

[0018] S23, a dynamic channel compression ratio vector β1 is introduced in each residual unit, and the number of compressed channels is defined as:

[0019]

[0020] Wherein, β l represents the channel compression ratio of the lth residual block, C l is the original output channel number of the residual block, C′ l is the number of compressed channels, represents rounding down;

[0021] S24, an asymmetric residual offset weight group is introduced in the residual connection structure, and the output of the lth residual unit is defined as:

[0022] y l =w m ·F l (x l )+w s ·x l +b l ;

[0023] Wherein, w m , w s respectively represent the response weights of the main branch and the shortcut connection path, F l (x l ) is the convolution output of the main branch, x l is the input feature map, b l is the bias term, and y l is the output of the residual unit;

[0024] S25, the output multi-layer feature maps C2, C3, C4 and C5 are input to the feature fusion module, a feature pyramid network structure is constructed, a top-down path fusion strategy is adopted, feature fusion is performed through upsampling and horizontal connection, and fused feature maps M2, M3 and M4 are obtained respectively;

[0025] S26, perform 3x3 convolution operation on the fusion feature maps M2, M3 and M4 respectively to obtain scale feature maps P2, P3 and P4 as the fingerprint ridge structure representation under different scales;

[0026] S27, introduce scale-aware fusion gating threshold τ s , for each scale feature map P i , calculate the response energy E(P i ) to determine whether to participate in the final fusion, and the fusion rule is:

[0027]

[0028] wherein, E(P i ) represents the response energy of the i-th scale feature map, τ s is the scale-aware gating threshold;

[0029] S28, the scale feature maps satisfying E(P i )≥τ s are spliced in the channel dimension to construct the final fusion feature map F;

[0030] S29, input the fusion feature map F into the classification module, and then input into the full connection layer group after flattening operation to generate the output vector Y;

[0031] S210, the output vector Y is used as the fingerprint identity recognition result for identity classification or similarity matching task.

[0032] Optionally, the S3 specifically comprises:

[0033] S31, initialize the population size N of the whale optimization algorithm, the maximum iteration number T max , the convergence control factor a, and randomly generate the initial population individual set:

[0034]

[0035] wherein, each whale individual W i represents a fingerprint recognition model parameter combination vector:

[0036] W i =[β2,β3,β4,β5,w m ,w s ,τ s ] i ;

[0037] S32, construct the fitness function , the model parameter compression rate R(W i ), and the model parameter compression rate R(W i), the multi-scale feature map information entropy H(F) is an evaluation index, and the fitness function is defined as follows:

[0038]

[0039] Wherein, λ1, λ2, λ3 are weighting coefficients;

[0040] S33, for each iteration process, according to the current optimal individual W * The distance D of the target individual W i Perform whale surrounding update mechanism:

[0041] W i (t+1) = W * (t) - A·|C·W * (t) - W i (t) |;

[0042] Wherein, r1, r2 are random numbers, t is the current iteration round, |·| represents the absolute value of the vector, W i (t+1) is the position of the i-th individual in the t+1 generation, W i (t) is the position of the i-th individual in the t generation, A is a convergence factor vector, C is a scaling factor vector, W * (t) is the parameter vector of the individual with the best fitness in the current population in the t generation.

[0043] S34, in each iteration, perform spiral position update strategy to simulate the hunting process of prey:

[0044] W 1 i (t+1) = D ′ ·e b·l ·cos*(2πl) + W * (t);

[0045] Wherein, W 1 i (t+1) is the updated position, b is a spiral shape constant, l is a random disturbance factor, D ′ represents the Euclidean distance between the current whale individual and the optimal individual, cos is the cosine function.

[0046] S35, select new individuals according to the update rule, calculate the fitness value If better than the previous generation result, update the current optimal individual W * (t+1);

[0047] S36, the iteration process continues until the termination condition is met, and the global optimal parameter combination W * is output.

[0048] S37, combine the optimal parameter combination W * Each parameter value in W is configured to each structure position of the fingerprint recognition model, wherein:

[0049] β l applied to residual unit channel compression control;

[0050] w m , w s applied to residual connection path weighting;

[0051] τ s applied to scale feature selection gate in feature fusion module;

[0052] S38, use the configured optimal parameter combination to complete model construction as the basic structure configuration of the training phase.

[0053] Optionally, the S4 specifically comprises:

[0054] S41, using the optimal parameter combination obtained by the whale optimization algorithm, the structure of the fingerprint recognition model is configured and initialized, including setting the channel compression ratio parameter of the residual network, the path weight parameter of the residual connection and the feature fusion gate threshold parameter;

[0055] S42, load the fingerprint recognition training data set, execute the standardization preprocessing process for each image, and construct a training batch;

[0056] S43, input the standardized fingerprint image into the configured fingerprint recognition model, execute the feature extraction, feature fusion and classification output process in sequence, and generate the prediction result of the fingerprint recognition model;

[0057] S44, construct a class boundary constraint type metric learning loss function for training, guide the fingerprint recognition model to output more discriminative embedding representation vectors by optimizing the intra-class compactness and inter-class separation degree;

[0058] S45, based on the loss function feedback result, update the weight parameters of each layer of the fingerprint recognition model using the back propagation algorithm, repeat the training round until the loss converges or meets the training termination condition;

[0059] S46, save the final fingerprint recognition model structure and parameter configuration after training, form a deployable and executable fingerprint recognition model, and use it for online fingerprint recognition tasks.

[0060] Optionally, the S5 specifically comprises:

[0061] S51, load the fingerprint recognition model which is trained and has complete parameter configuration to the target system, the target system comprising a fingerprint image acquisition device, an embedded processing module and an identification result output interface;

[0062] S52, real-time collection of user fingerprint images by a fingerprint image collection device, and standardization preprocessing of the collected images, including image alignment, normalization and enhancement operations, to generate standardized input images;

[0063] S53, inputting the standardized fingerprint images into a fingerprint recognition model, and sequentially performing convolution operations and residual propagation operations in the feature extraction module to extract a basic fingerprint line feature map;

[0064] S54, in the feature fusion module, a multi-scale feature map is fused and processed using a feature pyramid network structure to generate a fused multi-scale line feature map as a discriminant basis;

[0065] S55, inputting the fused feature map into a classification module, generating an identification vector through fingerprint recognition model inference, and combining the fingerprint recognition model output to calculate a fingerprint recognition result and a corresponding confidence score;

[0066] S56, transmitting the recognition result and the confidence score to the human-computer interface or the upper application module of the target system to support identity recognition verification, access control, authentication record, and complete the online execution process of the fingerprint recognition task.

[0067] The beneficial effects of the present application are:

[0068] The present application builds a deep fingerprint recognition model that integrates a dynamic channel compression mechanism, an asymmetric residual connection structure and a scale perception feature fusion strategy, and introduces a whale optimization algorithm to jointly optimize the model structure parameters and training parameters, achieving adaptive adjustment and recognition accuracy improvement of the fingerprint recognition system in complex image environments. Compared with the existing fixed structure recognition model, the improved method proposed in the present application not only effectively retains key line features and reduces redundant information interference, but also dynamically adjusts the model capacity and fusion path according to the actual characteristics of the fingerprint image, effectively improving the discrimination ability and resource utilization efficiency of the model.

[0069] In addition, the introduction of the whale optimization algorithm makes the model parameter configuration process free from the bottleneck of relying on manual debugging, and through the population intelligent search mechanism, the recognition accuracy is guaranteed while the computational complexity and training cost are greatly reduced. Especially in the model deployment stage, the present application can automatically generate the optimal structure scheme according to the characteristics of the target platform, with strong migratability and adaptability. Overall, the present application improves the stability and accuracy of the fingerprint recognition system in low-quality images, complex backgrounds, mobile terminals and other use environments, with comprehensive advantages such as high recognition rate, low resource consumption and strong generalization ability, and has good engineering application value and industrial promotion prospects. BRIEF DESCRIPTION OF DRAWINGS

[0070] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0071] Figure 1 A flow chart of a fingerprint recognition method based on deep learning proposed in the application;

[0072] Figure 2 A feature pyramid fusion structure based on scale perception gating mechanism of a fingerprint recognition method based on deep learning proposed in the application. DETAILED DESCRIPTION

[0073] The application will now be described in further detail with reference to the drawings. These drawings show only the essential features of the application and are therefore to be regarded only as a schematic representation of the basic structure of the application.

[0074] REFERENCE Figure 1 and Figure 2 A fingerprint recognition method based on deep learning, comprising the following steps:

[0075] S1, collecting fingerprint image data, pre-processing the fingerprint image data to obtain standardized fingerprint images;

[0076] S2, constructing a fingerprint recognition model, the fingerprint recognition model comprising a feature extraction module, a feature fusion module and a classification module, wherein the feature extraction module uses a residual neural network-based backbone network to extract the basic features of the fingerprint image, the feature fusion module introduces a feature pyramid network structure to up-sample and horizontally connect the basic features to generate multi-scale fusion feature maps, and the classification module is used to input the multi-scale fusion features into a fully connected layer to output the fingerprint identity recognition result;

[0077] S3, jointly optimizing the structure parameters and training parameters of the fingerprint recognition model based on a whale optimization algorithm to obtain an optimal parameter combination;

[0078] S4, configuring the fingerprint recognition model with the optimal parameter combination, inputting the standardized fingerprint images for training, optimizing the model weights using a class boundary-based metric learning loss function, and completing the fingerprint recognition model training process;

[0079] S5, deploying the trained fingerprint recognition model in a target system, receiving real-time collected fingerprint images, executing the image input, feature extraction, feature fusion and identity recognition processes, outputting the recognition result and confidence score, and realizing online execution of the fingerprint recognition task.

[0080] In this embodiment, the fingerprint image data specifically includes static grayscale images and dynamic infrared images.

[0081] In this embodiment, the pre-processing of the fingerprint image data specifically includes image grayscale normalization, noise filtering, geometric alignment and ridge enhancement operations.

[0082] In this embodiment, S2 specifically includes:

[0083] S21, construct the overall structure of the fingerprint recognition model, including a feature extraction module, a feature fusion module and a classification module, the input is a standardized fingerprint image, and the output is a fingerprint identity recognition result;

[0084] S22, in the feature extraction module, a residual neural network-based backbone structure is used to receive the standardized fingerprint image input, and a plurality of residual units are used to output a plurality of layers of basic feature maps C2, C3, C4 and C5 in turn, for representing fingerprint ridge features at different levels;

[0085] S23, a dynamic channel compression ratio vector β1 is introduced in each residual unit, and the number of compressed channels is defined as:

[0086]

[0087] wherein β l represents the channel compression ratio of the lth residual block, C l is the original output channel number of the residual block, C′ l is the number of compressed channels, represents the floor function;

[0088] S24, an asymmetric residual offset weight group is introduced in the residual connection structure, and the output of the lth residual unit is defined as:

[0089] y l = w m · F l (x l ) + w s · x l + b l ;

[0090] wherein w m , w s represent the response weights of the main branch and the shortcut connection path respectively, F l (x l ) is the convolution output of the main branch, x l is the input feature map, b l is the bias term, and y l is the output of the residual unit;

[0091] S25, input the output multi-layer feature maps C2, C3, C4 and C5 to a feature fusion module, construct a feature pyramid network structure, adopt a top-down path fusion strategy, perform feature fusion through upsampling and horizontal connection, and obtain fusion feature maps M2, M3 and M4 respectively;

[0092] S26, perform 3x3 convolution operations on the fusion feature maps M2, M3 and M4 respectively to obtain scale feature maps P2, P3 and P4 as representations of the fingerprint ridge structure at different scales;

[0093] S27, introduce a scale-aware fusion gating threshold τ s , for each scale feature map P i , calculate a response energy E(P i ) to determine whether to participate in the final fusion, and the fusion rule is:

[0094]

[0095] wherein E(P i ) represents the response energy of the i-th scale feature map, τ s is a scale-aware gating threshold;

[0096] S28, perform channel dimension splicing on the scale feature maps that satisfy E(P i ) >= τ s , and construct a final fusion feature map F;

[0097] S29, input the fusion feature map F to a classification module, perform a flattening operation, and then input to a fully connected layer group to generate an output vector Y;

[0098] S210, output the vector Y as a fingerprint identity recognition result for identity classification or similarity matching tasks.

[0099] In the embodiment, the S3 specifically includes:

[0100] S31, initialize the population size N of the whale optimization algorithm, the maximum number of iterations T max , the convergence control factor a, and randomly generate an initial population individual set:

[0101]

[0102] wherein each whale individual W i represents a fingerprint recognition model parameter combination vector:

[0103] W i = [β2, β3, β4, β5, w m , w s , τ s ] i ;

[0104] S32, constructing fitness function Validation set fingerprint identification accuracy A(W i ), model parameter compression rate R(W i ), multi-scale feature map information entropy H(F) as evaluation index, fitness function is defined as follows:

[0105]

[0106] Wherein, λ1, λ2, λ3 are weighting coefficients;

[0107] S33, for each iteration process, according to the current optimal individual W * And the distance D of the target individual W i Perform whale surrounding update mechanism:

[0108] W i (t+1) = W * (t)-A·|C·W · (t)-W i (t) |;

[0109] Wherein, r1, r2 are random numbers, t is the current iteration round, |·| represents the absolute value of vector, W i (t+1) is the position of the i-th individual in the t+1 generation, W i (t) is the position of the i-th individual in the t generation, A is the convergence factor vector, C is the scaling factor vector, W · (t) is the parameter vector of the individual with the optimal fitness in the current population in the t generation;

[0110] S34, in each iteration, perform spiral position update strategy to simulate the prey trapping process:

[0111] W 1 i (t+1) = D ′ ·e b·l ·cos(2πl)+W * (t);

[0112] Wherein, W 1 i (t+1) is the updated position, b is the spiral shape constant, l is a random disturbance factor, D ′ Indicates the Euclidean distance between the current whale individual and the optimal individual, cos is the cosine function;

[0113] S35, select new individuals according to the update rule, calculate the fitness value If better than the previous generation result, update the current optimal individual W *(t+1);

[0114] S36, the iteration process continues until the termination condition is met, and the global optimal parameter combination W is output * ;

[0115] S37, the optimal parameter combination W * Each parameter value is configured to each structure position of the fingerprint recognition model, wherein:

[0116] β l Applied to residual unit channel compression control;

[0117] w m , w s Applied to residual connection path weighting;

[0118] τ s Applied to scale feature selection gate in feature fusion module;

[0119] S38, using the configured optimal parameter combination to complete model construction as the basic structure configuration of the training stage.

[0120] In the embodiment, the S4 specifically comprises:

[0121] S41, using the optimal parameter combination obtained by the whale optimization algorithm to perform structure configuration and initialization of the fingerprint recognition model, including setting the channel compression ratio parameter of the residual network, the path weight parameter of the residual connection and the feature fusion gate threshold parameter;

[0122] S42, load the fingerprint recognition training data set, execute the standardization preprocessing process for each image, and construct the training batch;

[0123] S43, input the standardized fingerprint image into the configured fingerprint recognition model, execute the feature extraction, feature fusion and classification output process in turn, and generate the fingerprint recognition model prediction result;

[0124] S44, construct a class boundary constraint type metric learning loss function for training, guide the fingerprint recognition model to output more discriminative embedding representation vectors by optimizing the intra-class compactness and inter-class separation degree;

[0125] S45, based on the loss function feedback result, update the weight parameters of each layer of the fingerprint recognition model using the back propagation algorithm, repeat the training round until the loss converges or the training termination condition is met;

[0126] S46, save the final fingerprint recognition model structure and parameter configuration after training to form a deployable and executable fingerprint recognition model for online fingerprint recognition tasks.

[0127] In the embodiment, the S5 specifically comprises:

[0128] S51, load the fingerprint recognition model with completed training and parameter configuration to a target system, the target system including a fingerprint image acquisition device, an embedded processing module and an identification result output interface;

[0129] S52, acquire a user fingerprint image in real time through the fingerprint image acquisition device, and perform standardization preprocessing on the acquired image, including image alignment, normalization and enhancement operations, to generate a standardized input image;

[0130] S53, input the standardized fingerprint image into the fingerprint recognition model, and sequentially perform convolution operation and residual propagation operation in the feature extraction module to extract a basic fingerprint line feature map;

[0131] S54, in the feature fusion module, utilize a feature pyramid network structure to perform fusion processing on the multi-scale feature maps to generate a fused multi-scale line feature map as a discrimination basis;

[0132] S55, input the fused feature map into the classification module, generate an identification vector through fingerprint recognition model reasoning, and calculate a fingerprint recognition result and a corresponding confidence score in combination with fingerprint recognition model output;

[0133] S56, transmit the identification result and the confidence score to a man-machine interface or an upper application module of the target system, support identity recognition verification, access control, authentication record, and complete an online execution process of the fingerprint recognition task.

[0134] Embodiment 1:

[0135] In order to verify the feasibility of the present application in implementation, the platform collects and tests 10 kinds of typical scenes including standard environment, light interference, fingerprint pollution, offset and outdoor, 500 different user fingerprint images are sampled for each scene for comparative experiment, and the model proposed in the present application and a mainstream residual convolution fingerprint recognition model in an industry are used as comparative schemes.

[0136] In the present application scheme, the system is deployed with a residual dynamic channel compressor mechanism, an asymmetric residual connection structure, a scale perception FPN module, and the parameters β1, w m / w s , τ s are automatically searched and optimized by combining a whale optimization algorithm, and the model is finally deployed on an ARM architecture edge computing terminal. The average time consumption from fingerprint image acquisition to identification completion is less than 150 ms, and a higher identification accuracy is still maintained under test conditions with large image quality fluctuations (such as fingerprint pollution, low light and wet hands).

[0137] From the experimental results can be seen, in the "low light scene", the recognition accuracy of the present application scheme is 93.8%, and the comparative scheme is only 84.7%, which is improved by nearly 9 percentage points; in the "fingerprint pollution" case, the accuracy of the present application scheme reaches 89.4%, and the comparative method is only 76.5%, and the false recognition rate is reduced from 8.2% to 3.6%; in terms of response speed, the average response time of the present application method is always lower than 140ms, which is significantly better than the performance of more than 200ms of the comparative model.

[0138] From the actual use conditions of "dry hands", "wet hands", "mobile terminal" and the like, the present application model shows better robustness and adaptability, especially in the mobile terminal deployment test, the edge side is realized within 0.14 seconds. High reliability is completed to identify inference.

[0139] Table 1 performance comparison data of the present application method and the comparative scheme under different test scenarios

[0140]

[0141]

[0142] According to the data in Table 1, it can be seen that the fingerprint recognition method proposed by the present application is significantly better than the comparative scheme in various typical application environments, whether in recognition accuracy, response speed or false recognition rate and other key performance indicators, it has obvious advantages.

[0143] In the standard laboratory environment, the recognition accuracy of the present application method is 97.6%, which is improved by 6.3 percentage points compared with the comparative scheme, which shows that even in ideal conditions, the present application has stronger feature discrimination ability. The advantage is more obvious in the actual complex scene: the accuracy is improved from 84.7% to 93.8% in low light environment; in the case of fingerprint pollution, the improvement is greater, from 76.5% to 89.4%; in the offset and strong light interference conditions, it is also improved by 13.5% and 11.1% respectively, which fully shows that the present application can still maintain high robustness in poor image quality.

[0144] In terms of response time, the present application method is significantly better than the comparative scheme in all test scenarios. For example, in the fingerprint pollution scene, the average response time of the comparative scheme is 211ms, while the present application scheme is only 140ms, which is shortened by 71ms, effectively improving the response efficiency in actual application. In the real use environment of mobile terminal collection, dry hands, wet hands and the like, the response time of the present application method is stable at about 138ms, which is much lower than the more than 200ms of the traditional model.

[0145] Regarding the false recognition rate, the present invention's solution is generally more than 2% lower than the comparative solution. Especially in scenarios where false recognition is most likely to occur, such as "wet fingerprints" and "smudged fingerprints," the false recognition rate is reduced from 9.1% and 8.2% to 3.9% and 3.6%, respectively. This indicates that the introduced scale-aware fusion gating mechanism and dynamic channel compression strategy effectively reduce redundant information interference and improve the clarity of classification boundaries.

[0146] Based on the data shown in the table, this invention improves recognition accuracy while also considering response speed and false recognition control. The system performance is more stable and reliable, and it is suitable for online fingerprint recognition applications in various terminal devices and complex environments. It has strong engineering practicality and promotional value.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fingerprint recognition method based on deep learning, characterized in that, Includes the following steps: S1. Collect fingerprint image data, preprocess the fingerprint image data to obtain standardized fingerprint images; S2. Construct a fingerprint recognition model, which includes a feature extraction module, a feature fusion module, and a classification module. The feature extraction module uses a backbone network based on a residual neural network to extract the basic features of the fingerprint image. The feature fusion module introduces a feature pyramid network structure to upsample and laterally connect the basic features to generate a multi-scale fused feature map. The classification module is used to input the multi-scale fused features into a fully connected layer and output the fingerprint identity recognition result. S3. Based on the whale optimization algorithm, the structural parameters and training parameters of the fingerprint recognition model are jointly optimized to obtain the optimal parameter combination; S4. Configure the fingerprint recognition model using the optimal parameter combination, input a standardized fingerprint image for training, optimize the model weights using a metric learning loss function based on class boundaries, and complete the fingerprint recognition model training process. S5. Deploy the trained fingerprint recognition model in the target system, receive fingerprint images acquired in real time, execute the image input, feature extraction, feature fusion and identity recognition process, and output the recognition result and confidence score to realize the online execution of the fingerprint recognition task.

2. The fingerprint recognition method based on deep learning according to claim 1, characterized in that, The fingerprint image data specifically includes static grayscale images and dynamic infrared images.

3. The fingerprint recognition method based on deep learning according to claim 1, characterized in that, The preprocessing of fingerprint image data specifically includes image grayscale normalization, noise filtering, geometric alignment, and ridge enhancement operations.

4. The fingerprint recognition method based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Construct the overall structure of the fingerprint recognition model, including a feature extraction module, a feature fusion module, and a classification module. The input is a standardized fingerprint image, and the output is the fingerprint identity recognition result. S22. In the feature extraction module, a backbone structure based on a residual neural network is used to receive the standardized fingerprint image input, and multiple residual units sequentially output multi-layer basic feature maps C2, C3, C4 and C5 to characterize the fingerprint ridge features at different levels. S23. Introduce a dynamic channel compression ratio vector β1 in each residual unit, and define the number of channels after compression as follows: Where, β l C represents the channel compression ratio of the l-th residual block. l C′ represents the original output channel number of the residual block. l This represents the number of channels after compression. Indicates rounding down; S24. Introduce asymmetric residual offset weight reassembly in the residual connection structure, and define the output of the l-th residual unit as: y l =w m ·F l (x l )+w s ·x l +b l ; Among them, w m w s F represents the response weights of the main branch and the shortcut path, respectively. l (x l The main branch's convolution output, x l For the input feature map, b l For the bias term, y l This is the output of the residual unit; S25. Input the output multi-layer feature maps C2, C3, C4, and C5 into the feature fusion module to construct a feature pyramid network structure. Adopt a top-down path fusion strategy and perform feature fusion through upsampling and lateral connections to obtain fused feature maps M2, M3, and M4 respectively. S26. Perform 3×3 convolution operations on the fused feature maps M2, M3 and M4 respectively to obtain scale feature maps P2, P3 and P4, which serve as fingerprint ridge structure representations at different scales. S27. Introduce a scale-aware fusion gating threshold τ s For each scale feature map P i Calculate the response energy E(P) i To determine whether to participate in the final fusion, the fusion rules are as follows: Not participating in the integration; Among them, E(P) i ) represents the response energy of the i-th scale feature map, τ s Scale-aware gating threshold; S28, satisfying E(P) i )≥τ s The scale feature maps are concatenated along the channel dimension to construct the final fused feature map F; S29. Input the fused feature map F into the classification module, and after flattening, input it into the fully connected layer group to generate the output vector Y. S210, the output vector Y, serves as the fingerprint identification result, used for identity classification or similarity matching tasks.

5. The fingerprint recognition method based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Initialize the population size N for the whale optimization algorithm and the maximum number of iterations T. max Convergence control factor 'a', initial population set randomly generated: Among them, each individual whale W i This represents a vector of parameters for a fingerprint recognition model: W i =[β2,β3,β4,β5,w m ,w s ,t s ] i ; S32. Constructing the fitness function Validation set fingerprint recognition accuracy A(W) i Model parameter compression ratio R(W) i The multi-scale feature map information entropy H(F) is used as the evaluation index, and the fitness function is defined as follows: Where λ1, λ2, λ3 are weighting coefficients; S33. For each iteration process, based on the current optimal individual W * With target individual W i The distance D executes the whale encirclement update mechanism: W i (t+1)=W * (t)-A·|C·W * (t)-W i (t)|; Where r1 and r2 are random numbers, t is the current iteration round number, |·| represents the absolute value of the vector, and W i (t+1) represents the position of the i-th individual in generation t+1, W i (t) represents the position of the i-th individual in generation t, A is the convergence factor vector, C is the scaling factor vector, and W is the position of the ith individual in generation t. * (t) is the parameter vector of the individual with the best fitness in the current population in generation t; S34. In each iteration, execute the spiral position update strategy to simulate the prey encirclement process: W 1 i (t+1)=D′·e b·l ·cos(2πl)+W * (t); Among them, W 1 i (t+1) is the updated position, b is the spiral shape constant, l is the random perturbation factor, D′ represents the Euclidean distance between the current whale individual and the optimal individual, and cos is the cosine function; S35. Select a new individual according to the update rules and calculate the fitness value. If the result is better than the previous generation, then update the current best individual W. * (t+1); S36. The iterative process continues until the termination condition is met, and the globally optimal parameter combination W is output. * ; S37. Combine the optimal parameters W * The parameter values ​​in the configuration are respectively assigned to the structural positions of the fingerprint recognition model, where: β l Application in residual unit channel compression control; w m ,w s Applied to residual join path weighting; τ s Scale feature selection gating applied in the feature fusion module; S38. Use the optimal combination of configured parameters to complete the model construction, which serves as the basic structural configuration for the training phase.

6. The fingerprint recognition method based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. The optimal parameter combination obtained by the whale optimization algorithm is used to configure and initialize the fingerprint recognition model, including setting the channel compression ratio parameter of the residual network, the path weight parameter of the residual connection, and the feature fusion gating threshold parameter. S42. Load the fingerprint recognition training dataset, perform a standardized preprocessing procedure on each image, and build training batches; S43. Input the standardized fingerprint image into the configured fingerprint recognition model, and sequentially execute the feature extraction, feature fusion and classification output process to generate the fingerprint recognition model prediction result; S44. Construct a class boundary-constrained metric learning loss function for training. By optimizing intra-class compactness and inter-class separation, guide the fingerprint recognition model to output a more discriminative embedding representation vector. S45. Based on the feedback results of the loss function, the backpropagation algorithm is used to update the weight parameters of each layer of the fingerprint recognition model, and the training rounds are repeated until the loss converges or the training termination condition is met. S46. After training is completed, save the final fingerprint recognition model structure and parameter configuration to form a deployable fingerprint recognition model for online fingerprint recognition tasks.

7. The fingerprint recognition method based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Load the trained and fully configured fingerprint recognition model into the target system, wherein the target system includes a fingerprint image acquisition device, an embedded processing module, and a recognition result output interface; S52. Collect user fingerprint images in real time through fingerprint image acquisition device, and perform standardized preprocessing on the collected images, including image alignment, normalization and enhancement operations, to generate standardized input images; S53. Input the standardized fingerprint image into the fingerprint recognition model, and sequentially execute the convolution operation and residual propagation operation in the feature extraction module to extract the basic fingerprint ridge feature map. S54. In the feature fusion module, the feature pyramid network structure is used to fuse the multi-scale feature maps to generate a fused multi-scale ridge feature map as the basis for discrimination. S55. Input the fused feature map into the classification module, generate the recognition vector through fingerprint recognition model inference, and calculate the fingerprint recognition result and corresponding confidence score by combining the fingerprint recognition model output. S56. Transmit the recognition results and confidence scores to the target system's human-machine interface or upper-layer application module to support identity verification, access control, authentication records, and complete the online execution process of the fingerprint recognition task.

Images