A method of fingerprint comparison identification
By extracting fingerprint feature vectors using the FPRNet model and performing subtraction operations, the problems of low accuracy and efficiency in the recognition of noisy and low-quality fingerprint images in existing technologies are solved, and efficient fingerprint matching and recognition is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YUNNAN UNIV
- Filing Date
- 2023-04-17
- Publication Date
- 2026-05-01
AI Technical Summary
Existing fingerprint recognition technologies cannot effectively handle fingerprint images with high noise and poor image quality, resulting in low recognition accuracy and efficiency. They also have complex network structures and low ability to capture features from unprocessed fingerprint images.
A deep learning approach is adopted, which extracts fingerprint feature vectors through the FPRNet model, performs subtraction operations through the Subtract function, and finally determines fingerprint matching through a fully connected layer. This reduces network complexity, the number of parameters, and improves recognition accuracy and efficiency.
It improves the accuracy and efficiency of fingerprint recognition, reduces the complexity of the network, and enables efficient recognition of noisy and low-quality fingerprint images.
Smart Images
Figure CN116486440B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of identity security technology, and specifically to a method for fingerprint comparison and recognition. Background Technology
[0002] Fingerprints are an important biometric feature of the human body, widely used due to their lifelong immutability and uniqueness. In previous research, many scientists have proposed numerous algorithms and techniques in the field of fingerprint recognition. Simple sub-minimum matching is used to capture feature points and obtain information such as location and orientation. Then, various matching algorithms are used to calculate the fingerprint recognition rate. Two representative methods are: one is based on orientation field extraction and feature point matching. Fingerprint orientation field extraction mainly includes gradient-based methods, filtering-based methods, model-based algorithms, and neural network-based methods. Typical algorithms include: using different gradient direction ratios for fingerprint recognition, extracting the orientation field through matched filters, and extracting the orientation field using frequency domain filters. The other method is fingerprint feature point and singular point matching, with typical algorithms including LBP, HOG, and SIFT. However, these algorithms are susceptible to the effects of noise and fingerprint image quality.
[0003] Employing superior preprocessing techniques can significantly reduce the cost of subsequent network construction and training, enabling the rapid construction of custom deep learning networks. Currently, the approach focuses on three main aspects: fingerprint image enhancement, binarization techniques, and thinning techniques. Image enhancement includes spatial filtering, temporal filtering, and frequency filtering. Temporal filtering is fast at removing most noise from images, but its edge removal is poor. Frequency filtering offers good edge denoising, but its processing speed is relatively slow. Binarization techniques mainly include global thresholding, OTSU binarization, local thresholding, and edge operator methods. Global thresholding is fast, but the processed image quality is poor. Local thresholding is effective for abrupt changes in grayscale values. OTSU binarization focuses on the grayscale distribution of the image itself, and edge operator methods offer good edge processing. Thinning techniques mainly include non-iterative operator methods and iterative operator methods. Non-iterative operator methods are fast for image pruning, but the image thinning process is incomplete. Iterative operator methods solve the problem of incomplete images, but the computational load increases and the processing time becomes longer.
[0004] Therefore, existing fingerprint recognition technologies have the following drawbacks: they cannot handle fingerprint images with a lot of noise and poor image quality; the recognition accuracy and efficiency of unprocessed images are very low; and the existing fingerprint recognition network structure is complex and has a low ability to capture features of unprocessed fingerprint images, resulting in low fingerprint recognition rate and efficiency. Summary of the Invention
[0005] The purpose of this invention is to address the limitations of current fingerprint recognition technologies in handling noisy and poor-quality fingerprint images; the low accuracy and efficiency of recognition for unprocessed images; and the complex structure and limited ability of existing fingerprint recognition networks to extract features from unprocessed fingerprint images, resulting in low fingerprint recognition rates and efficiency. This invention provides a fingerprint comparison and recognition method that employs deep learning to extract fingerprint feature vectors from the fingerprint image via a network, then subtracts two feature vectors, and finally extracts features from the resulting vectors. The output indicates whether two fingerprints match, reducing the complexity of existing networks, decreasing the number of parameters, and improving fingerprint recognition accuracy and efficiency.
[0006] The technical solution of the present invention is as follows:
[0007] A fingerprint comparison and identification method includes the following steps:
[0008] Acquire paired fingerprint images to be compared, preprocess the fingerprint images, and divide the fingerprint images into training set and test set;
[0009] Building an FPRNet model involves three modules and the following steps:
[0010] Module 1 and Module 2 extract the feature vectors of the fingerprint image and compress the extracted feature vectors;
[0011] The Subtract function is used to subtract the feature vectors of Module 1 and Module 2 and then input them into Module 3;
[0012] Module 3 performs further feature extraction on the obtained feature vectors, and inputs the final feature vectors into a fully connected layer for comparison and recognition to determine whether the fingerprints match.
[0013] The preprocessed fingerprint images from the training set are input into the constructed FPRNet model for training, and the trained FPRNet model is obtained through iterative loops.
[0014] The pre-processed fingerprint images from the test set are input into the trained FPRNet model, and the recognition rate is obtained through iterative comparison.
[0015] Furthermore, module one and module two have the same network structure, both having Layer network model, the first The result of the layer output is The formula is: ,
[0016] in Indicates the first level to the second level. The tensors output by each layer, It is a composite function. It includes convolutional layers, batch normalization (BN) layers, pooling layers, and activation function operations.
[0017] Furthermore, the formula for the Subtract function is: ,in, This represents the feature vector output by module one. This represents the feature vector output by module two. This represents the input value for Module 3.
[0018] Furthermore, the feature vector obtained by module three The final one-dimensional vector is obtained through the module 3 network. The formula is: ,
[0019] in, It is a composite function, including convolutional layers, batch normalization (BN) layers, flattening layers, fully connected layers, and activation function operations.
[0020] Furthermore, the formula for determining whether a fingerprint matches is as follows:
[0021] ;
[0022] in, For the threshold, ,express The values at each position in the vector, when A value of 0 indicates that the two fingerprints are identical. A value of 1 indicates that the two fingerprints are not the same.
[0023] Furthermore, the determination of whether the fingerprint matches also includes a loss function, which is the binary cross-entropy loss function, and the formula is:
[0024] ;
[0025] .
[0026] Furthermore, the activation function uses Adam as the optimizer to optimize the output result. The gradient update rule formula for the Adam algorithm is:
[0027] ,
[0028] in Represents the gradient. express The momentum value at time t. express The exponentially decaying mean at time t.
[0029] Furthermore, the Batch Normalization (BN) layer is used to accelerate the convergence speed of the network. The specific steps are as follows: first, calculate the mean and variance of the samples; then, standardize the sample data; finally, perform translation and scaling processing, using the following formulas:
[0030] ;
[0031] ;
[0032] ;
[0033] in, The sample mean. For the sample variance, use and Perform an affine transformation to obtain the final BN output. .
[0034] Furthermore, the activation function is a ReLU activation function or a Sigmoid activation function.
[0035] Compared with existing technologies, the advantages of this invention are:
[0036] 1. A fingerprint comparison and recognition method, which uses deep learning to extract fingerprint feature vectors from fingerprint images through a network, then subtracts two feature vectors, and finally extracts features from the obtained feature vectors. The output result is whether the two fingerprints match. This reduces the complexity of existing networks, reduces the number of parameters, and improves fingerprint recognition accuracy and efficiency.
[0037] 2. A fingerprint comparison and recognition method that uses the ReLU function as the activation function. The ReLU function overcomes the gradient vanishing problem and speeds up the overall network training. Attached Figure Description
[0038] Figure 1 This is a flowchart of a fingerprint comparison and recognition method.
[0039] Figure 2 This is a flowchart of the FPRNet model structure, a fingerprint comparison and recognition method.
[0040] Figure 3 This is a network structure diagram of module one and module two of a fingerprint comparison and recognition method.
[0041] Figure 4 This is a network structure diagram of module three of a fingerprint comparison and recognition method.
[0042] Figure 5 This is a sample fingerprint dataset for a fingerprint comparison and recognition method.
[0043] Figure 6 This is an experimental result diagram of a fingerprint comparison and recognition method. Detailed Implementation
[0044] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0045] The features and performance of the present invention will be further described in detail below with reference to embodiments.
[0046] Please see Figure 1-6 A fingerprint comparison and identification method includes the following steps:
[0047] Acquire paired fingerprint images to be compared, preprocess the fingerprint images, and divide the fingerprint images into training set and test set; the preprocessing includes denoising, grayscale image processing, and binarization. The denoising method is Gaussian filtering, and the binarization method is global thresholding.
[0048] Building an FPRNet model involves three modules and the following steps:
[0049] Module 1 and Module 2 extract the feature vectors of the fingerprint image and compress the extracted feature vectors;
[0050] The Subtract function is used to subtract the feature vectors of Module 1 and Module 2 and then input them into Module 3;
[0051] Module 3 performs further feature extraction on the obtained feature vectors, and inputs the final feature vectors into a fully connected layer for comparison and recognition to determine whether the fingerprints match.
[0052] The preprocessed fingerprint images from the training set are input into the constructed FPRNet model for training, and the trained FPRNet model is obtained through iterative loops.
[0053] The pre-processed fingerprint images from the test set are input into the trained FPRNet model, and the recognition rate is obtained through iterative comparison.
[0054] Module 1 and Module 2 have the same network structure, both having Layer network model, the first The result of the layer output is The formula is: ,
[0055] in Indicates the first level to the second level. The tensors output by each layer, It is a composite function. It includes convolutional layers, batch normalization (BN) layers, pooling layers, and activation function operations.
[0056] Module 1 and Module 2 contain four hidden layers. The purpose is to obtain the feature vector of the output result from the network structure of Module 1 and Module 2 of the 240mm×320mm image. Then, the two vectors are subtracted and input into the next network part to finally obtain the feature vector to be identified. By comparison, it can be determined whether two fingerprints are similar.
[0057] The formula for the Subtract function is: ,in, This represents the feature vector output by module one. This represents the feature vector output by module two. This represents the input value for Module 3.
[0058] The feature vector obtained by module three The final one-dimensional vector is obtained through the module 3 network. The formula is: ,
[0059] in, It is a composite function, including convolutional layers, batch normalization (BN) layers, flattening layers, fully connected layers, and activation function operations.
[0060] Since Module 1 and Module 2 have the same network structure, when the same fingerprint image is input, the corresponding values in the output feature vector exhibit a certain proportional relationship. Therefore, a threshold T is set so that when two fingerprint images are identical, the values in the feature vector after the subtraction operation are equal to... Most of the values in the fingerprint feature vectors obtained by subtracting from different fingerprint images are not equal to... The obtained feature vector is then used to extract features, and the final output feature vector is passed through a fully connected layer to determine whether two fingerprints are the same. The formula for determining whether fingerprints match is as follows:
[0061] ;
[0062] in, For the threshold, ,express The values at each position in the vector, when A value of 0 indicates that the two fingerprints are identical. A value of 1 indicates that the two fingerprints are not the same.
[0063] The process of determining whether a fingerprint matches also includes a loss function, which is the binary cross-entropy loss function, and the formula is:
[0064] ;
[0065] .
[0066] For the loss function between probabilities, when and When they are equal, Only if it is 0, otherwise It's a positive number. And the greater the difference in probability, the better. The larger it is.
[0067] The activation function uses Adam as the optimizer to optimize the output result. The gradient update rule formula of the Adam algorithm is as follows:
[0068] ,
[0069] in Represents the gradient. express The momentum value at time t. express The exponentially decaying mean at time t.
[0070] The Batch Normalization (BN) layer is used to accelerate the convergence speed of the network. The specific steps are as follows: first, calculate the mean and variance of the samples; then, standardize the sample data; finally, perform translation and scaling operations, using the following formulas:
[0071] ;
[0072] ;
[0073] ;
[0074] in, The sample mean. For the sample variance, use and Perform an affine transformation to obtain the final BN output. .
[0075] The activation function is either ReLU or Sigmoid. Modules one and two use ReLU as the activation function, which overcomes the vanishing gradient problem and speeds up the overall network training. Adjusting the images to the same size through the input layer facilitates feature point extraction. Module three uses both ReLU and Sigmoid activation functions.
[0076] Experimental Analysis
[0077] Environment setup
[0078] The hardware environment used in the experiment was as follows: CPU: Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz + 1.99GHz; GPU: NVIDIA GeForce MX250; RAM: 16GB. The software deep learning framework used was TensorFlow version 2.6.0, CUDA version 11.2, and the data packages used included Keras, os, sys, and argparse. The operating system used was Windows 10 64-bit. The programming language was Python 3.9.
[0079] Dataset
[0080] The dataset consists of fingerprint images from the FVC2000 and FVC2002 datasets, with the remaining portion collected by a touch-sensitive fingerprint scanner. Fingerprints were collected from 20 individuals, with 10 fingerprints per person. Images were then captured from different orientations for each fingerprint type, with five orientations for each fingerprint. The collected images were then rotated and cropped. The majority of the self-collected fingerprints were blurry, damaged, or incomplete. The final dataset comprises approximately 7500 fingerprint images. Examples of the data are provided below. Figure 5 As shown.
[0081] Experimental steps
[0082] It consists of three steps:
[0083] (1) For a size of 240mm The 320mm image is preprocessed by fingerprint image enhancement, grayscale image processing, and binarization to obtain the fingerprint image required for the training of this network. The preprocessed image is then divided into training set and test set.
[0084] (2) The processed fingerprint images are fed into the network for training. Through continuous iteration, a network model with good training performance is finally obtained. As the number of iterations increases, the loss decreases continuously. Therefore, choosing an appropriate number of iterations can not only shorten the network training time, but also obtain a fairly good training result. Based on the experimental results, the number of iterations was selected as 2500.
[0085] (3) The fingerprints of the test set are also preprocessed and then fed into the trained network. The recognition rate is obtained through iterative comparison.
[0086] Analysis of Experimental Results
[0087] After 2500 iterations, the final network model was obtained, and the training results were as follows: Figure 6 As shown in the figure, the FPRNet network model achieved a fingerprint recognition rate of 98.72%. Then, a test set was used to verify the model's accuracy and robustness. The ratio of the training set to the test set was 9:1, therefore the test set contained 640 fingerprint images. These test set images were then processed through the FPRNet network model, and the final test results are shown in Table 1.
[0088] Table 1: Recognition rate of fingerprint images on the test set
[0089]
[0090] The experimental results show that the overall fingerprint image recognition rate was 98.75%, which is a good result.
[0091] To verify that the designed FPRNet model has a high recognition rate on unprocessed data, experiments were conducted using unprocessed fingerprints. The experimental dataset used was FVC2000. The experimental results are shown in Table 2.
[0092] Table 2: Recognition rates obtained by deep neural networks for images with and without preprocessing
[0093]
[0094] The FPRNet network, after being designed, achieved a fingerprint recognition rate of 98.23% without preprocessing the fingerprints, demonstrating that the FPRNet network has a high recognition rate for unclear or blurry fingerprint images.
[0095] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.
Claims
1. A method for fingerprint comparison and recognition, characterized in that, Includes the following steps: Acquire paired fingerprint images to be compared, preprocess the fingerprint images, and divide the fingerprint images into training set and test set; Building an FPRNet model involves the following steps: The FPRNet model consists of three modules: Module 1 and Module 2 extract feature vectors from the fingerprint image and compress the extracted feature vectors. The Subtract function is used to subtract the feature vectors of Module 1 and Module 2 and then input them into Module 3; Module 3 performs further feature extraction on the obtained feature vectors, and inputs the final feature vectors into a fully connected layer for comparison and recognition to determine whether the fingerprints match. The preprocessed fingerprint images from the training set are input into the constructed FPRNet model for training, and the trained FPRNet model is obtained through iterative loops. The pre-processed fingerprint images from the test set are input into the trained FPRNet model, and the recognition rate is obtained through iterative comparison. Module 1 and Module 2 have the same network structure, both having Layer network model, the first The result of the layer output is The formula is: , in Indicates the first level to the second level. The tensors output by each layer, It is a composite function. Includes convolutional layers, batch normalization (BN) layers, pooling layers, and activation function operations; The formula for the Subtract function is: ,in, This represents the feature vector output by module one. This represents the feature vector output by module two. This represents the input value for Module 3; Module 3 will obtain the feature vector The final one-dimensional vector is obtained through the module 3 network. The formula is: , in, It is a composite function, including convolutional layers, batch normalization (BN) layers, flattening layers, fully connected layers, and activation function operations.
2. The fingerprint comparison and recognition method according to claim 1, characterized in that, The formula for determining whether a fingerprint matches is as follows: ； in, For the threshold, ,express The values at each position in the vector, when A value of 0 indicates that the two fingerprints are identical. A value of 1 indicates that the two fingerprints are not the same.
3. The fingerprint comparison and recognition method according to claim 1, characterized in that, The process of determining whether a fingerprint matches also includes a loss function, which is the binary cross-entropy loss function, and the formula is: ； 。 4. The fingerprint comparison and recognition method according to claim 1, characterized in that, The activation function uses Adam as the optimizer to optimize the output result. The gradient update rule formula of the Adam algorithm is as follows: ， in Represents the gradient. express The momentum value at time t. express The exponentially decaying mean at time t.
5. The fingerprint comparison and recognition method according to claim 1, characterized in that, The Batch Normalization (BN) layer is used to accelerate the convergence speed of the network. The specific steps are as follows: first, calculate the mean and variance of the samples; then, standardize the sample data; finally, perform translation and scaling operations, using the following formulas: ； ； ； in, The sample mean. For the sample variance, use and Perform an affine transformation to obtain the final BN layer output. .
6. The fingerprint comparison and recognition method according to claim 1, characterized in that, The activation function is either a ReLU activation function or a Sigmoid activation function.