Palmprint crease simulation curve-based fine palmprint sample generation method

Abstract

The invention discloses a fine palmprint sample generation method based on a palmprint crease simulation curve, and the method comprises the following steps: S1, a generation mechanism of the palmprint crease simulation curve is invented, so that the generated simulation curve better accords with the crease trend and distribution of a real palmprint; s2, converting the real palmprint into a corresponding curve graph with a crease simulation curve style by using a conversion module of the invention; s3, performing fine tuning by using a pre-trained vector quantization variational auto-encoder, and mapping a real palm print to a potential representation with a lower dimension; s4, matching a curve graph with a crease simulation curve style with the potential representation of the real palmprint, and performing matched data training of the diffusion model; and S5, generating a required number of palm print crease simulation curves, and inputting the palm print crease simulation curves into a model to sample a fine palm print sample. According to the method, the problem of low palmprint generation quality at present is solved, a more vivid palmprint picture can be generated, people can use the palmprint picture to train the recognition model under the condition that real palmprint data is deficient, and the recognition precision of the recognition model is improved.

Description

Technical Field

[0001] The invention belongs to the field of palmprint feature extraction and generation in biological feature extraction and generation, and in particular relates to the generation of fine pseudo palmprints using palmprint crease simulation curves. Background Art

[0002] Palmprint recognition is not as widely used as fingerprint recognition and face recognition due to the limitation of data loss. Existing palmprint generation models are basically based on generative adversarial networks in deep learning. There is still a certain gap between the generated details and realism and the real palmprints. The generated palmprints cannot completely replace the real palmprints to train the recognition model. At the same time, the controllability of the palmprint generation model is also a major problem. Using the palmprint crease simulation curve as a control condition requires that the palmprint crease simulation curve can well reflect the crease trend and distribution of the real palmprint. The real palmprint is affected by the lighting and sampling equipment in the actual sampling process. The palmprints obtained by sampling have various styles, and there are certain rotations and displacements, which also brings huge challenges to palmprint generation. Generating realistic and diverse pseudo-palmprints, expanding the training data of the training recognition model, and thus improving the recognition accuracy of the recognition model has become an important research and development direction. Summary of the invention

[0003] The present invention aims to solve the deficiencies of the prior art and proposes a method for generating fine palmprint samples based on a palmprint crease simulation curve. The diffusion model can be used to generate high-quality realistic images, the vector quantization autoencoder can reduce the size of the diffusion space, and the invented palmprint crease simulation curve can well reflect the real palmprint crease, etc., so that realistic and fine palmprint samples can be generated.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A method for generating a fine palmprint sample based on a palmprint crease simulation curve comprises the following steps:

[0006] S1. Using the generation mechanism of palmprint crease simulation curve, the generated simulation curve is made more consistent with the crease trend and distribution of the real palmprint, and a sufficient number of palmprint crease simulation curves B are generated, which can serve as the control condition input in the generation stage S5, and can also serve as the conversion domain of the palmprint crease simulation pseudo-curve conversion module in S2, guiding the model to generate a pseudo-curve C with the style of palmprint crease simulation curve B and maintaining the crease distribution of the real palmprint A;

[0007] S2. converting the real palm print A into a corresponding curve graph C in the style of palm print crease simulation curve B;

[0008] S3. Use the pre-trained vector quantized variational autoencoder for fine-tuning to map the real palmprint A to a lower-dimensional latent representation L, and enable the decoder of the vector quantized variational autoencoder to remap the latent representation L back to a reconstructed image R that is almost the same as the real palmprint A;

[0009] S4. Pair the pseudo curve C with the crease simulation curve B grid and the latent representation L of the real palmprint to perform matching data training of the diffusion model, and set the number of steps T of gradually adding noise to the latent representation L. T should be large enough to ensure that when the number of steps is T, the latent representation L added with noise is approximately a white noise graph of the standard normal distribution;

[0010] S5. Use the required number of palmprint crease simulation curves B generated by S1 to input the model to sample fine palmprint samples.

[0011] Preferably, the S1 comprises: establishing a palmprint crease simulation curve generation module, the function of which is to generate a required number of palmprint crease simulation curves B by performing random sampling and parameter adjustment based on the 12 palmprint crease simulation curve templates designed by the invention.

[0012] Preferably, S2 includes: establishing a palmprint crease simulation pseudo curve conversion module, the function of which is to extract the palmprint crease of the real palmprint A as a grayscale binary curve graph P, and then train the generator to perform style transfer to convert the grayscale binary curve graph P into a pseudo curve C with the style of the palmprint crease simulation curve B.

[0013] Preferably, S3 includes: establishing a real palmprint latent space mapping module, the function of which is to map the real palmprint A to a latent space with a lower dimension to obtain a latent representation L. Due to the lower dimension, the training time can be efficiently saved during the diffusion model training (i.e. S4), and the reasoning time of the generation stage (i.e. S5) can also be significantly reduced.

[0014] Preferably, S4 includes: establishing a matching data diffusion training module, the function of which is to use the potential representation L (obtained in S3) and the pseudo curve C (obtained in S2) that match the real palm print A to establish matching data, train the diffusion model, and establish a connection between the potential representation L and the standard normal distribution white noise.

[0015] Preferably, S5 includes: establishing a fine palmprint sample generation module, the function of which is to use the palmprint crease simulation curve generation module established in S1 to generate a required number of palmprint crease simulation curves B as input. The palmprint crease simulation curve B controls the crease distribution and trend of the generated palmprint sample, and the generated palmprint sample and the palmprint crease simulation curve B have a high degree of identity consistency.

[0016] The present invention takes advantage of the diffusion model that can generate high-quality images, invents an improved palm print crease simulation curve, and creatively uses the invented palm print crease simulation pseudo curve conversion module to convert the real palm print into a pseudo curve with the style of the palm print crease simulation curve, so as to obtain a matching pseudo curve with a one-to-one correspondence with the real palm print. In the process of using matching data to train the diffusion model, the pseudo curve acts as a control condition, so that in the generation stage, the palm print crease simulation curve generated on demand can be used as a model input to generate high-quality and fine palm print samples. The sample can be used to train the recognition model. In the case of a lack of real palm print samples, the recognition accuracy of the recognition model is greatly improved, which is conducive to the promotion and deployment of palm print recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0018] Figure 1 12 template images of palm print crease simulation curves according to an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of palmprint crease simulation curve generation according to an embodiment of the present invention;

[0020] Figure 3 This is a palmprint crease simulation curve generation effect diagram of an embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of model training according to an embodiment of the present invention;

[0022] Figure 5 This is a schematic diagram of the effect of generating a pseudo-curve of a palmprint crease according to an embodiment of the present invention;

[0023] Figure 6 A schematic diagram of a palmprint sample collection process according to an embodiment of the present invention;

[0024] Figure 7 A diagram showing the relationship between the modules in an embodiment of the present invention;

[0025] Figure 8 This is a diagram showing the effect of palmprint sampling in 10 steps according to an embodiment of the present invention.

[0026] Fig. 9 This is a 20-step palmprint sample sampling effect diagram of an embodiment of the present invention;

[0027] Fig.10 This is a 1000-step palmprint sample sampling effect diagram of an embodiment of the present invention;

[0028] Fig.11 This is a comparison diagram of the sampling effects of different numbers of palmprint samples according to an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0030] First, some nouns or terms that appear in the description of the embodiments of the present application are subject to the following interpretation:

[0031] VQ-VAE: The full name is Vector Quantized Variational Autoencoder. VQ-VAE is a generative model based on autoencoder. VQ-VAE discretizes data representation, enabling it to process image, speech and other data more efficiently. This discretization is achieved by introducing vector quantization, which allows the model to avoid the limitations of continuous distribution when learning potential representations, thereby obtaining clearer results in generation tasks.

[0032] DDPM: The full name is Denoising Diffusion Probabilistic Model. DDPM is a generative model based on the diffusion process. It gradually adds noise to the data to form a forward process from data distribution to Gaussian noise. Then a neural network is trained to reversely denoise and gradually restore the original data distribution from the noise. This method performs well in tasks such as image generation, especially in terms of image quality and details, which are significantly improved compared with previous generative models.

[0033] DDIM: The full name is Denoising Diffusion Implicit Model. DDIM is an improved diffusion model derived from the original DDPM model. It proposes a new sampling method that can reduce the sampling steps while maintaining the generation quality, thereby increasing the generation speed. Unlike the Markov chain used by DDPM, DDIM uses a deterministic non-Markov process, allowing the model to generate high-quality images in fewer steps. Compared with DDPM, DDIM can reduce the degradation of generated image quality under the same accelerated sampling premise. Therefore, it is widely used in the generation process of accelerated diffusion models while maintaining a high generation effect.

[0034] CUT: The full name is Contrastive Unpaired Translation. CUT is an image-to-image translation method without paired data, proposed by researchers at the University of California, Berkeley. Unlike traditional image translation methods (such as CycleGAN) that rely on cycle consistency loss, CUT uses contrastive learning to achieve inter-domain mapping, and completes the unsupervised image translation task by maximizing the similarity of images of the same type and the difference of images of different types in the latent space. CUT is more computationally efficient and can also maintain good performance in image quality.

[0035] U-Net: It is a convolutional neural network architecture that was originally designed for biomedical image segmentation tasks and later widely used in various image segmentation tasks. The name U-Net comes from the "U" shape of its network structure, which is symmetrically composed of an encoder and a decoder.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] The detailed description of S1 and the palmprint crease simulation curve generation module established in this step is as follows:

[0038] like Figure 2 As shown in the palm print crease simulation curve generation process, first set a parameter point boundary so that the synthesized palm print crease simulation curve is within the square plane area. The parameter points mentioned in the present invention include the starting point, the control point and the end point. The main line M mentioned in the present invention is the Mth main line from top to bottom. Further, the present invention defines 3 sets of main line parameter points, each set of parameter points can form 4 main line templates, and these main line templates are different. Among them, the main line template formed by parameter point 1 is as follows Figure 1 As shown in row (a), its characteristics are: optionally, the starting points of main line 2 and main line 3 are set to coincide with each other with a probability of 0.50; add a fourth main line, optionally, add a main line from the lower right to the upper left with a probability of 0.70, or add a main line from the lower left to the upper right with a probability of 0.30. The main line template formed by parameter point 2 is as follows Figure 1 As shown in row (b), Figure 1 The difference of the template shown in row (a) is that its main line 2 has two control points. Optionally, control point 1 makes the front half of the main line 2 bend upward, and control point 2 makes the back half of the main line 2 bend downward. The main line template formed by parameter point 3 is as follows: Figure 1As shown in row (c), its characteristic is that, considering the situation of broken palms in the crowd, the number of main lines of these main line templates is 2 to 3. Further, a set of main line parameter points is selected according to probability, and then the main line template is determined. Optionally, parameter point 1 is selected with a probability of 0.50, parameter point 2 is selected with a probability of 0.47, and parameter point 3 is selected with a probability of 0.03. Further, the direction and length of the fine lines are set. In order to make the generated fine lines more consistent with muscle texture, for the direction of the fine lines, it is optional to be located at (0, π / 6) with a probability of 0.15, at (5π / 6, 7π / 6) with a probability of 0.2, at (11π / 6, 2π) with a probability of 0.15, at (π / 3, 2π / 3) with a probability of 0.2, at (4π / 3, 5π / 3) with a probability of 0.2, and at other positions of (0, 2π) with a probability of 0.1. For the length of fine lines, optionally, the probability of 0.80 is located at (50-100) and the probability of 0.20 is located at (100-150) to increase diversity. Further, a simulated palm print crease curve graph containing main lines and fine lines is synthesized and binarized. Optionally, the pixel value 128 is set as the threshold. If the pixel value is greater than 128, it is white, and if it is less than 128, it is black. Finally, the processed simulated palm print crease curve graph is saved. Figure 3 As shown, the simulated palmprint crease curve diagram generated by the present invention can find similar palmprint creases in the real palmprint database, and is generated according to different probabilities, with inter-class diversity.

[0039] The detailed description of S2 and the palmprint crease simulation pseudo-curve conversion module established in this step is as follows:

[0040] like Figure 4As shown in the palm print crease simulation pseudo curve conversion module in the lower left corner, the (size optional, description omitted later) 256×256 pixel color real palm print image A is first converted into its corresponding (size optional, description omitted later) 256×256 binary grayscale palm print crease extraction image P, and then the CUT training mode is used to train the generator G with several downsampling and upsampling blocks, and the palm print crease simulation curve B generated by the palm print crease simulation curve generation module in S1 is used as the style conversion domain of the palm print crease extraction image P, which is converted into a pseudo curve C with the style of the palm print crease simulation curve B. Since the output of the generator G is not a binary grayscale image, in order to ensure that the pseudo curve C is binary like the palm print crease simulation curve B, the feature map output by the generator G needs to be binarized. Optionally, the values ​​greater than 0 are set to 1, and those less than 0 are set to -1. Optionally, the corresponding real image can be normalized in Pytorch to set the pixel value between -1 and 1. Optionally, the crease extractor can be acted as by various low-frequency filters; the generator G can be acted as by neural network models such as U-Net, Resnet, etc. with several downsampling and upsampling. It is only necessary to ensure that the input and output have the same size. After the training of the generator G is completed, the palm print crease simulation pseudo-curve conversion module is established, which can convert the real palm print A into a pseudo-curve C that is specially used to reflect its crease trend and distribution and has the same style as the palm print crease simulation curve b. The pseudo-curve C has a one-to-one correspondence with the real palm print A, and the connection between the palm print crease simulation curve B and the real palm print A is established. Figure 5 By comparison, column (a) shows a randomly selected real palmprint A, column (b) shows the crease image P extracted from the real palmprint image using the palmprint crease simulation pseudo-curve conversion module, column (c) shows the pseudo-curve C generated by converting the real image through the palmprint crease simulation pseudo-curve conversion module, and column (d) shows the pseudo-curve generated directly from the real image through CUT. It can be observed that the pseudo-curve generated directly using CUT often misses some important folds and introduces additional unnecessary folds. However, the pseudo-curve generated using the palmprint crease simulation pseudo-curve conversion module largely retains the main folds of the original palmprint while avoiding the generation of additional interference curves.

[0041] The detailed description of S3 and the real palmprint latent space mapping module established in this step is as follows:

[0042] like Figure 4As shown in the real palmprint latent space mapping module above, the 256×256 real palmprint A is used as input and reduced in dimension through the VQ-VAE encoder, and mapped to the (optional size, omitted below) 64×64 three-channel feature map latent space to obtain the latent representation L of the real palmprint, reducing the representation dimension of the image, which can save the training time of the diffusion model and the inference time of the palmprint generation stage. Then, the latent representation L is vectorized and discretized, and then output as the reconstructed palmprint image R through the VQ-VAE decoder.

[0043] The VQ-VAE model pre-trained on other color datasets can be used for fine-tuning, and the total loss of fine-tuning is:

[0044] L VQ-VAE =L reconstruction +L VQ

[0045] in,

[0046] L reconstruction =||AR|| 1

[0047] L VQ =||sg[z e (A)-e]|| 2 +β||z e (A)-sg[e]|| 2

[0048] sg[.] means stop gradient operator, z e (A) is the encoder continuous latent representation of input A (i.e., L), and e is the e (A) The nearest quantized embedding vector. β is a hyperparameter that controls the weight of the loss.

[0049] The detailed description of S4 and the matching data diffusion training module established in this step is as follows:

[0050] like Figure 4 As shown in the matching data diffusion training module in the lower right corner, through the palmprint crease simulation pseudo-curve conversion module established in S2, we can obtain the pseudo-curve C corresponding to the real palmprint A, which reflects its crease trend and distribution. At the same time, the real palmprint latent space mapping module established in S3 is used to obtain the latent representation L of the real palmprint A. First, the pseudo-curve C is reduced to the same dimension as the latent representation L (optional 64×64), and then the diffusion step T of the latent representation L is set to 1000 (optional), and normal noise is gradually added to the latent representation L and the reduced pseudo-curve C is spliced ​​in the channel dimension as the input of the model U-Net. The training process uses the standard DDPM training method, in which the reduced pseudo-curve C serves as a control condition.

[0051] The detailed description of S5 and the matching data diffusion training module established in this step is as follows:

[0052] like Figure 6 In the fine palmprint sample generation module shown, pseudo curve C is no longer needed in the generation stage. A sufficient number of palmprint crease simulation curves B are generated by the palmprint crease simulation curve generation module in S1, and their resolution is reduced to 64×64 pixels, and they are spliced ​​together along the channel dimension with three-channel random noise sampled from a standard normal distribution. Palmprint crease simulation curve B is used as a conditional control to determine the crease of the palmprint, while the noise affects the details and style of the palmprint. The spliced ​​feature map is then input into the U-Net model in the matching data diffusion training module established in S4, and the latent representation of the latent space is obtained by sampling through DDIM or DDPM. Finally, the VQ-VAE decoder in the real palmprint latent space mapping module established in S2 is used to map the latent representation to a fine-grained pseudo palmprint image.

[0053] like Figure 7 As shown, the five modules in the five steps have the following connections:

[0054] S1: Palmprint crease simulation curve generation module

[0055] This module generates a simulated curve of palmprint creases to represent the shape of the wrinkles in the palmprint. The simulated curve B of the palmprint creases generated by the module is used as the input of the generator G to provide the style conversion area. At the same time, this module uses the simulated palmprint crease curve B as the control input to control the crease features in the generated palmprint sample.

[0056] S2: Palmprint crease simulation pseudo-curve conversion module

[0057] This module is responsible for converting the real palm print A into a pseudo curve C so as to match the crease pattern of the real palm print A and provide the pseudo curve C in the matching data for S4 to establish a matching data diffusion training module.

[0058] S3: Real palmprint latent space mapping module

[0059] This module can generate the latent space representation L of the real palmprint A, which is the matching data required by S4 to establish the matching data diffusion training module. At the same time, the decoder in the module will be used in the fine palmprint sample generation module S5.

[0060] S4: Matching data diffusion training module

[0061] This module provides matching data from S2 and S3 for diffusion process training.

[0062] S5: Fine palmprint sample generation module

[0063] This is the final module of the entire process. It generates a network, uses the palmprint crease curve B of S1 as input, gradually reduces noise through the U-Net model of S4, and finally generates a palmprint sample through the decoder of S3.

[0064] This process step works through the collaborative work of multiple modules, from simulating palmprint crease curves to latent space mapping and diffusion training, and then to detail generation, and finally outputs high-quality pseudo palmprint samples.

[0065] Figure 8 , 9 , 10 are sampling sample examples under different sampling steps when the time step T=1000, which are DDIM10-step sampling, DDIM20-step sampling, and DDPM1000-step sampling respectively. Fig.11 The effect display and comparison of different sampling steps of DDIM (s=10 means DDIM10 sampling, ResizedCondition is the scaled palmprint crease simulation curve) are shown in Figure 2. The present invention can generate high-quality palmprint samples with clear details, and the palmprint styles are diverse, and the crease trend and distribution are controllable. Theoretically, the present invention can generate palmprint samples of any number and different identities, so that the number of palmprint samples can be expanded, the recognition model can be trained, and the recognition model accuracy can be improved.

[0066] The following describes an optional embodiment of using a method for generating fine palmprint samples based on a palmprint crease simulation curve of the present invention to generate a palmprint sample training recognition model.

[0067] A total of 37,390 palm print images from 8 public data sets are used to extract the effective area in the middle of the palm as the real palm print A. The 8 public data sets are from: Indian Institute of Technology Delhi (IITD), Maintenance Planning Document (MPD), Trinity College Dublin (TCD), Sapienza University Mobile Palmprint Database (SMPD), Birjand University Mobile Palmprint Database (BMPD), College of Engineering, Pune (COEP), Mutah University Mobile Hand Images Database (MOHI), Digital Signal Processing Group Palmprint Database (GPDS).

[0068] Data name Number of identities Number of palm prints MPD 400 16000 TCD 600 12000 IITD 460 2601 BMPD 81 1604 SMPD 91 1871 COEP 157 1300 MOHI 128 1300 GPDS 99 714 sum 2016 37390

[0069] 37390 real palmprints are used as the real palmprint A of the present invention to establish a palmprint crease simulation curve generation module, a palmprint crease simulation pseudo-curve conversion module, a real palmprint latent space mapping module, a matching data diffusion training module, and a fine palmprint sample generation module. Optionally, a data set of 4000 identities, 100 palmprint samples for each identity, and a total of 400,000 palmprint samples are generated to train the recognition model. The real palmprint A is divided into a training data set and a verification data set with a ratio of 1006 to 1010 according to the identity. The recognition accuracy of the recognition model trained with the generated 400,000 palmprint samples and the recognition model trained with the real palmprint data are compared. The comparison index is TAR@FAR, and the trained recognition models are Resnet50 and MobileFaceNet. The following table shows the recognition accuracy on the BMPD and COEP datasets. Method A represents the recognition model result trained with the training set of 1006 identities of real palmprint A, method 20 represents the recognition model result trained with the DDIM20-step sampling dataset of the present invention, and method 10 represents the recognition model result trained with the DDIM10-step sampling dataset of the present invention. It can be analyzed that the palmprint samples generated by the present invention can significantly improve the recognition accuracy of the recognition model.

[0070]

[0071] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating fine palmprint samples based on palmprint crease simulation curves, characterized in that: The following steps are involved: S1. Invent a generation mechanism for palm print crease simulation curves, so that the generated simulation curves are more consistent with the crease trend and distribution of real palm prints; S2. Using the conversion module of the invention, the real palm print is converted into a corresponding curve graph with a crease simulation curve style; S3. Fine-tune the pre-trained vector quantized variational autoencoder to map the real palm print to a lower-dimensional latent representation; S4. Use the curve graph with crease simulation curve style and the potential representation of the real palm print to pair and perform matching data training on the diffusion model; S5. Generate the required number of palmprint crease simulation curves and input them into the model to sample fine palmprint samples.

2. A method for generating fine palmprint samples based on palmprint crease simulation curves according to claim 1, characterized in that: The step S1 includes: using real palm print data, analyzing the distribution and trend of palm print creases, and constructing a palm print crease simulation curve.

3. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 1, characterized in that: The S2 includes: Construct a conversion module from real palm print A to pseudo curve C, where pseudo curve C has the style of simulated palm print crease curve B. First, use the palm print crease extractor to extract the crease P of the real palm print, and then train the generator G to convert the crease P into pseudo curve C through style transfer training, where the style of pseudo curve C is consistent with that of simulated palm print crease curve B.

4. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 1, characterized in that: The S3 includes: By fine-tuning the pre-trained vector quantized variational autoencoder, the real palmprint A is mapped to a lower-dimensional latent representation L, reducing the representation dimension of the image, which can save the training time of the diffusion model and the inference time of the palmprint generation stage.

5. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 1, characterized in that: The S4 includes: Using the pseudo curve C and the latent representation L corresponding to the real palmprint A, the pixels of the pseudo curve C are first scaled to the same dimension as the latent representation L using the bilinear scaling method, and then their matching properties are used to train the diffusion model of the matching data.

6. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 1, characterized in that: The S5 includes: Generate the required number of simulated palmprint crease curves B, scale the pixels using the bilinear scaling method, and then randomly generate three-channel standard normal distribution random noise, which are spliced ​​in the channel dimension and input into the diffusion model to obtain the potential representation of the pseudo palmprint. Finally, the potential representation is passed through the decoder of the vector quantized variational autoencoder to generate a palmprint sample.

7. A method for generating fine palmprint samples based on palmprint crease simulation curves, the method applying the method according to any one of claims 1 to 6, characterized in that: include: Palmprint crease simulation curve generation module, palmprint crease simulation pseudo-curve conversion module, real palmprint latent space mapping module, matching data diffusion training module, and fine palmprint sample generation module.

8. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 7, characterized in that: The palmprint crease simulation curve generation module is used to generate an improved palmprint crease simulation curve B, which is used for training the palmprint crease simulation pseudo-curve conversion module and serves as the conversion domain of the style conversion model. At the same time, it serves as a control condition in the fine palmprint sample generation module as the input of the generation model to generate fine palmprint samples.

9. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 7, characterized in that: The palmprint crease simulation pseudo-curve conversion module converts the real palmprint A into a grayscale binary image C, which has the style of the palmprint crease simulation curve B, so that the palmprint crease simulation curve B can be directly used as input in the fine palmprint sample generation module based on the model trained by it. At the same time, the module obtains matching data corresponding to the real palmprint A, and can perform matching data diffusion model training in the matching data diffusion training module, which serves as a control condition.

10. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 7, characterized in that: The real palmprint latent space mapping module maps the real palmprint A to a latent space with a lower dimension, obtains a latent space representation L with a lower dimension than the real palmprint, and reduces the diffusion model training time and palmprint sample sampling time.

11. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 7, characterized in that: The matching data diffusion training module aims to gradually add normal noise to the latent representation L of the real palmprint A, and finally map it to the standard normal noise N of the same dimension, so as to establish the connection between the latent space and the standard normal distribution. When generating, the randomly generated noise N and the palmprint crease simulation curve B can be spliced ​​as the generation model input.

12. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 7, characterized in that: The purpose of the fine palmprint sample generation module is to use the palmprint crease simulation curve B generated in advance by the palmprint crease simulation curve generation module and the randomly generated standard normal noise N as the model input, and then splice the denoised noise and the palmprint crease simulation curve B into the model, and after T steps of iteration, the generated palmprint potential representation L' can be obtained, and finally the decoder of the vector quantization autoencoder in the real palmprint latent space mapping module is used to map the generated palmprint potential representation L' to the generated palmprint. The randomly generated noise N controls the details and style of the generated pseudo palmprint, and the palmprint crease simulation curve B controls the crease trend and direction of the generated pseudo palmprint.

13. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 8, characterized in that The palm print crease simulation curve generation module also includes referring to real palm print samples, simulating the diversity of real palm print samples, forming 12 real palm print crease simulation curve templates, sampling and modifying the 12 templates, so as to generate various palm print crease simulation curves.

14. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 9, characterized in that: The palmprint crease simulation pseudo-curve conversion module also includes a palmprint crease extraction step, which uses a texture extractor to extract the real palmprint into a grayscale binary palmprint crease map P, and then uses the generative adversarial training of style transfer to train the generator to convert the palmprint crease map P into a palmprint crease simulation pseudo-curve C that is also grayscale binary.

15. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 10, characterized in that: The real palmprint latent space mapping module also includes fine-tuning using a pre-trained vector quantized variational autoencoder to obtain a latent representation L of the latent space of the real palmprint A, while the decoder can accurately restore the real palmprint A through the latent representation L.

16. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 11, characterized in that: The matching data diffusion training module also includes scaling the palm print crease simulation curve B by pixels to a dimension consistent with the real palm print potential representation L.

17. The method for generating fine palmprint samples based on palmprint crease simulation curve according to claim 12, characterized in that: The fine palmprint sample generation module also includes a module that can use the generated palmprint samples to train a recognition model, expand the sample training volume of the recognition model, and improve the recognition accuracy of the recognition model.

Images