Single-sample palm vein recognition method

Abstract

The invention belongs to the technical field of palm vein recognition, and particularly relates to a single-sample palm vein recognition method. Comprising the following steps: step 1, training a recurrent neural network-diffusion model used for generating a supplementary training sample by using a single palm vein image training sample of each user and a multi-scale and multi-direction transformation result of the training sample; 2, jointly forming a training sample set by using a single palm vein image training sample of a user and a supplementary training sample generated by the recurrent neural network-diffusion model based on the training sample, and training a palm vein image recognition network according to the training sample set; and step 3, recognizing a palm vein image by using the palm vein image recognition network trained in the step 2. According to the method, each user only needs one training image to train an identification model with relatively high identification efficiency, so that the work of collecting training samples is reduced, and the efficiency is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of palm vein recognition, and in particular relates to a single-sample palm vein recognition method. Background Art

[0002] Palm vein recognition technology is a biometric identification technology. After the palm is irradiated with near-infrared light, the palm capillaries have strong absorption of infrared rays, so darker lines will be left in the veins. Since the palm veins have characteristics such as papillary lines, main lines and wrinkles, each person's palm vein image is unique. Therefore, in today's information development, palm veins are used for biometric identification, used to identify users and protect personal information security. Palm vein recognition technology uses infrared waves to irradiate the palm to obtain a vein image, and then processes the image to obtain biometrics. Finally, the biometrics are matched with the database to determine the user's identity. Palm veins have great advantages over other biometric identification technologies. For example, palm veins are not affected by light, expression, posture, makeup, etc. At the same time, palm veins can only be imaged by specific wavelengths of light when the blood flows, and a clear image of palm veins can be obtained, which is difficult to forge, which improves the level of information security.

[0003] Existing palm vein recognition technology requires more samples as training data to train the model in order to achieve a high recognition accuracy. However, obtaining a large amount of training data is time-consuming and laborious, and requires a high degree of cooperation from users, which makes it difficult to obtain training data. Therefore, this method is inefficient. Summary of the invention

[0004] The present invention discloses a single-sample palm vein recognition method for palm vein recognition. In this method, each user only needs one training image to train a recognition model with high recognition efficiency, thereby reducing the work of collecting training samples and improving efficiency.

[0005] The single-sample palm vein recognition method of the present invention comprises:

[0006] Step 1, using a single palm vein image training sample of each user and the multi-scale and multi-directional transformation results of the training sample, training a recurrent neural network-diffusion model for generating supplementary training samples;

[0007] Step 2, forming a training sample set with a single palm vein image training sample of the user and a supplementary training sample generated by the recurrent neural network-diffusion model based on the training sample, and using the training sample set to train the palm vein image recognition network;

[0008] Step 3: Use the palm vein image recognition network trained in step 2 to recognize the palm vein image.

[0009] Furthermore, in step 1, in addition to the original image, the multi-scale and multi-directional transformation results include an image of the original image rotated 90 degrees, an image of the original image reduced by 1 / 2, and an image of the original image reduced by 1 / 2 + rotated 90 degrees.

[0010] Furthermore, the recurrent neural network-diffusion model is composed of a multi-scale, multi-directional transformation module and multiple diffusion models and recurrent neural networks;

[0011] A single palm vein image training sample is transformed through a multi-scale and multi-directional transformation module to obtain four different images: the original image, the original image rotated 90 degrees, the original image reduced by 1 / 2, and the original image reduced by 1 / 2 + rotated 90 degrees; the four images are respectively input into diffusion model 1, diffusion model 2, diffusion model 3 and diffusion model 4, each diffusion model consists of two parts: an interconnected coding and noise module and a decoding and noise removal module; adjacent coding and noise removal modules and adjacent decoding and noise removal modules are connected by a recurrent neural network;

[0012] The original image is input into the encoding and noise module of diffusion model 1, and the encoding result is input into the encoding and noise module of diffusion model 2 together with the original image rotated 90 degrees after passing through the recurrent neural network;

[0013] The encoding result of the diffusion model 2 encoding and noise module passes through the recurrent neural network and is input into the diffusion model 3 encoding and noise module together with the original image reduced by 1 / 2;

[0014] The encoding result of the encoding and noise module of diffusion model 3 passes through the recurrent neural network and is input into the encoding and noise module of diffusion model 4 together with the original image that is reduced by 1 / 2 and rotated by 90 degrees.

[0015] The encoding result of the diffusion model 4 encoding and noise module is directly input into the diffusion model 4 decoder denoising module. After the decoding result passes through the recurrent neural network, it is input into the diffusion model 3 decoder denoising module together with the encoding result of the diffusion model 3 encoding and noise module.

[0016] The decoding result of the diffusion model 3 decoder denoising module passes through the recurrent neural network and is then input into the diffusion model 2 decoder denoising module together with the encoding result of the diffusion model 2 encoding and noise module;

[0017] The decoding result of the diffusion model 2 decoder denoising module passes through the recurrent neural network and is then input into the diffusion model 1 decoder denoising module together with the encoding result of the diffusion model 1 encoding and noise module;

[0018] The data image is finally generated by the decoding and denoising module output of the diffusion model 1.

[0019] The present invention can generate high-quality data through the proposed recurrent neural network-diffusion model, thereby increasing training samples, and then in the training process of the palm vein image recognition network, each user only needs to provide one training sample to train a model with higher recognition accuracy; the model proposed by the present invention can reduce the demand for training data, thereby simplifying the collection of user data, improving the overall recognition rate, and improving the efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Flow chart of a single-sample palm vein recognition method in an embodiment of the present invention.

[0021] Figure 2 4 is a structural diagram of a recurrent neural network-diffusion model in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1 As shown in the process, the single-sample palm vein recognition method disclosed in this example is mainly divided into three stages: training the recurrent neural network (Recurrent Neural Network)-diffusion model (Diffusion Model), training the recognition network, and identifying samples. More high-quality data is generated by training the recurrent neural network-diffusion model, and added to the training samples as supplementary data for training the palm vein image recognition network. In the recognition sample stage, the sample to be classified is input into the recognition network, and the recognition network gives the recognition result.

[0023] Through this method, a better recognition model can be trained through one training sample for each user, thereby improving work efficiency and saving labor costs.

[0024] The structure diagram of the recurrent neural network-diffusion model in this example is as follows Figure 2 As shown, it is composed of a multi-scale, multi-directional transformation module, multiple diffusion models and a recurrent neural network; the training data image is transformed through the multi-scale, multi-directional transformation module to obtain four different images: the original image, the original image rotated 90 degrees, the original image reduced by 1 / 2, and the original image reduced by 1 / 2 + rotated 90 degrees. The above four images are respectively input into diffusion model 1, diffusion model 2, diffusion model 3 and diffusion model 4. Each diffusion model consists of two parts: an interconnected coding and noise module and a decoding and noise removal module; adjacent coding and noise removal modules and adjacent decoding and noise removal modules are connected by a recurrent neural network.

[0025] Specifically, the original image is input into the encoding and noise module of diffusion model 1, and the encoding result is input into the encoding and noise module of diffusion model 2 together with the original image rotated 90 degrees after passing through the recurrent neural network;

[0026] The encoding result of the diffusion model 2 encoding and noise module passes through the recurrent neural network and is input into the diffusion model 3 encoding and noise module together with the original image reduced by 1 / 2;

[0027] The encoding result of the encoding and noise module of diffusion model 3 passes through the recurrent neural network and is input into the encoding and noise module of diffusion model 4 together with the original image that is reduced by 1 / 2 and rotated by 90 degrees.

[0028] The encoding result of the diffusion model 4 encoding and noise module is directly input into the diffusion model 4 decoder denoising module. After the decoding result passes through the recurrent neural network, it is input into the diffusion model 3 decoder denoising module together with the encoding result of the diffusion model 3 encoding and noise module.

[0029] The decoding result of the diffusion model 3 decoder denoising module passes through the recurrent neural network and is then input into the diffusion model 2 decoder denoising module together with the encoding result of the diffusion model 2 encoding and noise module;

[0030] The decoding result of the diffusion model 2 decoder denoising module passes through the recurrent neural network and is then input into the diffusion model 1 decoder denoising module together with the encoding result of the diffusion model 1 encoding and noise module;

[0031] The data image is finally generated by the decoding and denoising module output of the diffusion model 1.

[0032] In this example, more specific parameters and network settings are well known to those skilled in the art and will not be described in detail here. For further details of the diffusion model, refer to the document "ASurvey on Generative Diffusion Models, HanqunCao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and StanZ.Li," (arXiv:2209.02646).

[0033] The training process of the RNN-Diffusion model is divided into the following steps:

[0034] 1) First, initialize the recurrent neural network and diffusion model, and initially set the network parameters;

[0035] 2) Perform multi-scale and multi-directional transformation on the training data to obtain images of different scales and directions;

[0036] 3) Train the recurrent neural network-diffusion model. By comparing the differences between the training images and the generated images, the network parameters of the recurrent neural network-diffusion are updated and adjusted.

[0037] Repeat the above three steps until the model parameters are optimal. This process generally needs to be repeated 1000 times.

[0038] After completing the training of the recurrent neural network-diffusion model, the recognition network is trained.

[0039] The training process of the recognition network is divided into the following steps:

[0040] 1) Initialize the recognition network, set the network structure and randomly generate network parameters.

[0041] 2) The data generated by the recurrent neural network-diffusion model and the training data are input into the recognition network together, the recognition network is trained, and the recognition error is obtained by a closed test method, and then the parameters of the recognition network are adjusted using this error.

[0042] Repeat the above steps 1000 times to complete the training of the recognition network.

[0043] In the recognition stage, the sample to be recognized is input into the trained recognition network, and the recognition network gives the recognition result.

Claims

1. A single sample palm vein recognition method, characterized in that: include: Step 1, using a single palm vein image training sample of each user and the multi-scale and multi-directional transformation results of the training sample, training a recurrent neural network-diffusion model for generating supplementary training samples; Step 2, forming a training sample set with a single palm vein image training sample of the user and a supplementary training sample generated by the recurrent neural network-diffusion model based on the training sample, and using the training sample set to train the palm vein image recognition network; Step 3: Use the palm vein image recognition network trained in step 2 to recognize the palm vein image.

2. The method according to claim 1, characterized in that: In step 1, in addition to the original image, the multi-scale and multi-directional transformation results include an image of the original image rotated 90 degrees, an image of the original image reduced by 1 / 2, and an image of the original image reduced by 1 / 2 + rotated 90 degrees.

3. The method according to claim 2, characterized in that The recurrent neural network-diffusion model is composed of multi-scale and multi-directional transformation modules as well as multiple diffusion models and recurrent neural networks; A single palm vein image training sample is transformed through a multi-scale and multi-directional transformation module to obtain four different images: the original image, the original image rotated 90 degrees, the original image reduced by 1 / 2, and the original image reduced by 1 / 2 + rotated 90 degrees; the four images are respectively input into diffusion model 1, diffusion model 2, diffusion model 3 and diffusion model 4, each diffusion model consists of two parts: an interconnected coding and noise module and a decoding and noise removal module; adjacent coding and noise removal modules and adjacent decoding and noise removal modules are connected by a recurrent neural network; The original image is input into the encoding and noise module of diffusion model 1, and the encoding result is input into the encoding and noise module of diffusion model 2 together with the original image rotated 90 degrees after passing through the recurrent neural network; The encoding result of the diffusion model 2 encoding and noise module passes through the recurrent neural network and is input into the diffusion model 3 encoding and noise module together with the original image reduced by 1 / 2; The encoding result of the encoding and noise module of diffusion model 3 passes through the recurrent neural network and is input into the encoding and noise module of diffusion model 4 together with the original image that is reduced by 1 / 2 and rotated by 90 degrees. The encoding result of the diffusion model 4 encoding and noise module is directly input into the diffusion model 4 decoder denoising module. After the decoding result passes through the recurrent neural network, it is input into the diffusion model 3 decoder denoising module together with the encoding result of the diffusion model 3 encoding and noise module. The decoding result of the diffusion model 3 decoder denoising module passes through the recurrent neural network and is then input into the diffusion model 2 decoder denoising module together with the encoding result of the diffusion model 2 encoding and noise module; The decoding result of the diffusion model 2 decoder denoising module passes through the recurrent neural network and is then input into the diffusion model 1 decoder denoising module together with the encoding result of the diffusion model 1 encoding and noise module; The data image is finally generated by the decoding and denoising module output of the diffusion model 1.

Images