A fundus image matching method, system and readable medium based on deep learning
Through the deep learning-based fundus image matching method, the deep neural network model and semi-supervised training scheme are used to solve the problem of difficult to achieve high performance in image registration and authentication in the prior art, and efficient image matching and authentication are achieved.
Patent Information
- Application Number
- CN202210667546.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-14
- Publication Date
- 2025-05-16
- Estimated Expiration
- 2042-06-14
AI Technical Summary
Existing fundus image matching technology is difficult to achieve high performance in multiple tasks such as image registration and authentication at the same time, due to limited labeling data and complex manual design features.
The fundus image matching method based on deep learning is adopted to realize key point detection and feature extraction through deep neural network models, and the encoder network and decoder network are used to extract features, and the calculation amount and labeled data requirements are reduced through semi-supervised training schemes.
It realizes high performance of image registration and authentication, can process larger images, is suitable for medical fundus image scenarios, and solves the problem of insufficient labeling data through PKE technology.
Smart Images
Figure CN114926892B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a fundus image matching method, system and readable medium, belonging to the technical field of image matching. Background Art
[0002] Retinal image matching technology aims to automatically determine the matching degree of two given fundus images based on the image content. The definition of the matching degree depends on the specific task. In the retina image registration task, the matching of two images depends on the calculation of the affine transformation of the corresponding pixels between the two. In the retinal image based identity verification task, the two images should be considered matched if and only if they come from the same eye.
[0003] Previous fundus image matching technologies were limited by limited annotated data, and were usually implemented using traditional digital image processing technologies and relied on complex artificially designed features. Artificial features often require strong prior knowledge when they are designed, and the matching features required by the two different tasks of image registration and identity verification are not consistent, so existing technologies can usually only be applied to a specific task. For example, an existing technology for identity verification adopts a strategy of constructing a matching template using the spatial features of vascular bifurcation points. In order to speed up identity verification, each image only generates a feature vector after extracting the principal component, so this technology cannot be used for image registration. In another existing technology, camera intrinsic parameters and spherical models are used to align the feature points of the fundus image in 3D space, but the speed and efficiency are far from meeting the requirements of identity verification. Therefore, it is difficult for traditional technologies to achieve high performance in multiple tasks such as image registration and identity verification at the same time. Summary of the invention
[0004] In response to the above problems, the purpose of the present invention is to provide a fundus image matching method, system and readable medium based on deep learning, which is simple and effective, and only needs to train part of the features instead of all the feature maps, which greatly reduces the amount of calculation and allows the input of larger images, and is suitable for processing fundus images with higher resolution.
[0005] To achieve the above-mentioned purpose, the present invention proposes the following technical scheme: a fundus image matching method based on deep learning, comprising: inputting the fundus image into a deep neural network model for training to obtain a deep neural network model that simultaneously realizes key point detection and feature extraction, the deep neural network model includes an encoder network and two decoder networks; inputting the fundus image to be tested into the encoder network of the trained deep neural network model to extract features, and then inputting the features into the two encoder networks respectively to generate a key point probability map and a description feature map; obtaining a key point coordinate set and a key point feature set according to the key point probability map and the description feature map; matching the key point coordinate set and the key point feature set corresponding to the two fundus images to be matched to obtain a matching set; calculating the parameters of the affine transformation according to the matching set to realize image registration; determining the matching outer points and inner points according to the matching set, and performing identity authentication according to the number of inner points, and if the number exceeds a specific threshold, the verification is passed.
[0006] Furthermore, the encoder network is responsible for extracting fundus image features, and the two decoder networks include a key point detection decoder and a feature description decoder. The key point detection decoder is responsible for completing image key point detection; the feature description decoder is responsible for completing key point feature description.
[0007] Furthermore, only part of the fundus image needs to carry key point annotation information, and the fundus image is input into the deep fundus image model. If the fundus image carries key point annotation information, the key point annotation information is input into the key point detection decoder as Y0.
[0008] Furthermore, the method for generating the key point probability map is as follows: by upsampling the high-level features in the key point detection decoder to obtain the same height and width dimensions as the low-level features in the key point detection decoder, the high-level features are spliced with the low-level features along the channel dimension, and then the spliced features are convolved and upsampled to obtain an output image with the same size as the input image and a channel number of 1, and the key point probability map is generated after sigmoid activation.
[0009] Furthermore, the description feature map is directly obtained by using the quadratic interpolation method, and its size is consistent with the original image.
[0010] Furthermore, the key point detection decoder uses a progressive key point amplification technique to extract key points, and the extracted key points do not include key point annotation information.
[0011] Furthermore, the key points are extracted using a two-stage filtering method: the first stage is geometry-based screening, which requires that the key points appearing in the fundus image should also appear in the transformation map; the second stage is content-based screening, which requires that the key point features detected in the fundus image and the key point features in the transformation map satisfy the optimal KNN matching.
[0012] Furthermore, the feature description decoder is based on the key point set obtained by the key point detection decoder Sampling is performed on the description feature map to obtain the feature set, and then based on the random affine transformation matrix Construct a triplet loss function for training.
[0013] Furthermore, the matching set uses the RANSAC algorithm or the median method to calculate the matching outer points and inner points, and the number of remaining inner points is used as an indicator of identity authentication. If the indicator is less than a preset threshold, the identity authentication is rejected, otherwise the identity authentication is accepted; or the RANSAC algorithm or the median method is used to directly solve the affine matrix that describes the changes in the fundus image, and image alignment is obtained through the affine matrix.
[0014] The present invention also discloses a fundus image matching system based on deep learning, which includes the following four modules:
[0015] A model training module, used for inputting the fundus image into a deep neural network model for training to obtain a deep neural network model that simultaneously realizes key point detection and feature extraction. The deep fundus image model includes an encoder network and two decoder networks;
[0016] A model output module is used to input the fundus image to be tested into a deep neural network model that simultaneously realizes key point detection and feature extraction, and generates a key point probability map and a description feature map;
[0017] A feature set extraction module is used to obtain a key point coordinate set and a key point feature set according to a key point probability map and description features;
[0018] The matching set generation module is used to match the key point coordinate set and the key point feature set corresponding to the two fundus images to be matched to obtain a matching set; the identity authentication module is used to calculate the parameters of the affine transformation according to the matching set to achieve image registration, determine the matching external points and internal points according to the matching set, and perform identity authentication based on the number of internal points.
[0019] The present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform a fundus image matching method based on any of the above-mentioned deep learning methods.
[0020] The present invention adopts the above technical solution, which has the following advantages:
[0021] 1. The solution of the present invention simply and effectively realizes image registration and identity verification at the same time.
[0022] 2. The present invention only needs to train some features instead of all feature maps, which greatly reduces the amount of calculation, allows the input of larger images, and is more suitable for medical fundus image scenarios.
[0023] 3. The present invention proposes a PKE technology for training a key point detection model, which gradually expands key points with high repeatability and high reliability in a semi-supervised manner, solves the problem of insufficient manually labeled data, and effectively enhances the performance of image matching.
[0024] 4. The present invention designs a deep fundus image model, which is more suitable for fundus image matching and has better performance than previous feature descriptors. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 is a flow chart of a fundus image matching method based on deep learning in one embodiment of the present invention;
[0026] Figure 2 is a structural diagram of a neural network model in an embodiment of the present invention. The number under each feature graph represents the output size of the feature graph, where h and w are the height and width of the input image respectively, and the d value is usually 256;
[0027] Figure 3 is a training flow chart of a key point detection decoder in one embodiment of the present invention;
[0028] Figure 4 It is a schematic diagram of key point screening using the progressive key point amplification technology in one embodiment of the present invention. Hollow points represent unmatched points that need to be filtered. DETAILED DESCRIPTION
[0029] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail through specific embodiments. However, it should be understood that the provision of specific embodiments is only for a better understanding of the present invention, and they should not be understood as limitations of the present invention. In the description of the present invention, it should be understood that the terms used are only for the purpose of description and cannot be understood as indicating or implying relative importance.
[0030] In order to solve the problem that it is difficult to achieve high performance in multiple tasks such as image registration and identity authentication in the prior art, the present invention proposes a fundus image matching method, system and readable medium, which adopts the technical route of deep learning to achieve fundus image matching. However, deep learning technology relies on a large amount of labeled data for training, while fundus medical matching data is scarce, and manual labeling is time-consuming and difficult to label completely. Therefore, it adopts a coding-double decoding neural network structure that is more suitable for medical fundus images, and combines a semi-supervised training scheme. Only a small amount of incomplete labels are needed to achieve efficient training of the model, solving the problem of shortage of labeled data. In real clinical data tests, the present invention has achieved the current best performance in both identity recognition and registration tasks. The scheme of the present invention is described in detail below with reference to the accompanying drawings through embodiments.
[0031] Embodiment 1
[0032] This embodiment discloses a fundus image matching method based on deep learning, such as Figure 1 As shown, including:
[0033] S1 inputs the fundus image into the deep fundus image model for training to obtain a trained deep fundus image model, where the deep fundus image model includes an encoder network and two decoder networks;
[0034] In order to avoid too much prior knowledge introduced by artificially designed features, this embodiment proposes a deep learning model specifically for fundus image matching, namely a deep fundus image model, which extracts matching features through gradient optimization of back propagation. The structure diagram and convolution settings of the deep fundus image model are as follows: Figure 2 As shown in the figure, the deep fundus image model includes a shared encoder network and two decoder networks at the same level. The encoder network is responsible for extracting fundus image features, and the two decoder networks include a key point detection decoder and a feature description decoder. The key point detection decoder is responsible for completing image key point detection; the feature description decoder is responsible for completing key point feature description.
[0035] Prepare fundus image data. Only a part of the fundus image needs to have key point annotation information. Input the fundus image into the deep fundus image model. If the fundus image has key point annotation information, input the key point annotation information into the key point detection decoder as Y0. Otherwise, no additional input is required. After setting the number of iterations and optimizer parameters, the deep fundus image model is automatically optimized through iterative back propagation without manual intervention.
[0036] S2 inputs the fundus image to be tested into the trained deep fundus image model to generate a key point probability map and a description feature map;
[0037] In the key point detection decoder network, since many key feature patterns in the fundus image are relatively low-level, such as vascular feature patterns, and too deep a network layer will cause the extracted high-level features to lose the low-level feature details of the fundus image, which is not conducive to finding vascular bifurcations. Therefore, in this embodiment, the key point probability map is obtained by short-circuiting the low-level features and the high-level features to retain the low-level feature parameters and extract the key point features; the description feature map is directly obtained by the quadratic interpolation method, and its size is consistent with the original image.
[0038] The following symbols can be defined to describe the scheme in the embodiment:
[0039]
[0040] Among them, I represents the input image of the model, F represents the abstract feature obtained by the encoder, P is the key point probability map obtained by the detection decoder, and D represents the description feature map obtained by the description decoder. Represents the set of key points obtained by performing non-maximum suppression (NMS) on P.
[0041] S3 obtains a key point set and a sub-feature set according to the key point probability map and the description features;
[0042] As shown in the above formula, non-maximum suppression is performed on the key point probability map to obtain a key point set. Then, based on the key point set, sampling is performed on the corresponding pixel position of the description feature map to obtain the corresponding descriptor feature set.
[0043] In this embodiment, the problem that fundus medical images are poorly labeled and cannot be trained on deep models is solved by using a semi-supervised training method. The specific contents are as follows:
[0044] The key point detection decoder uses Progressive Keypoint Expansion (PKE) technology to extract key points, and the extracted key points do not include key point annotation information.
[0045] like Figure 3 As shown, randomly generate an affine transformation matrix Apply this matrix to image I The transformation graph I′ is obtained, which can be expressed as Let I′ and I pass through the same encoding-keypoint detection decoder to obtain the corresponding probability maps P′ and P, then the loss function L geo for DICE loss between P and
[0046]
[0047] Where i,j represents the (i,j)th pixel of the feature map, represents matrix dot multiplication. The loss term aims to improve the stability of key point detection under different viewing angles. In addition, the loss function L clf is the label Y after iterative expansion based on the initial input label Y0 t The DICE loss between is used to introduce supervisory information. t The process is as Figure 3 As shown, input I′ and I, the key point detection decoder detects the corresponding key point set, and among these key points there are some key points that are not included in Y0 but are still beneficial to matching. The goal of PKE is to retain these key points and make these key points further guide the training of the detector.
[0048] In this embodiment, the key points are extracted using a two-stage filtering method:
[0049] The first stage is geometry-based screening, such as Figure 4 As shown by the cross at the bottom of the figure, the key points that appear in the fundus image should also appear in the transformation map; the second stage is content-based screening, such as Figure 4 As shown by the second to last cross in the figure, the key point features detected in the fundus image and the key point features in the transformation graph are required to meet the optimal KNN matching, that is, the minimum distance is less than the second smallest distance multiplied by the decimal parameter. The first stage can be regarded as the screening of the repeatability of the key points, and the second stage is the screening of the reliability of the key points. The key points that pass the two screenings can be added to Y t Finally, Gaussian filtering is performed on the candidate key point set to obtain soft labels for supervised training of detection descriptors.
[0050] The feature description decoder uses triplet loss for unsupervised training. The key point set obtained by the key point detection decoder Sampling is performed on the description feature map to obtain the feature set, and then based on the random affine transformation matrix A triplet loss function is constructed for training. This scheme is simple and effective, and only requires training some features instead of all feature maps, which greatly reduces the amount of calculation and allows input of larger images, making it more suitable for medical fundus image scenarios.
[0051] S4 matches the key point set and the sub-feature set to obtain a matching set;
[0052] The KNN matching strategy is used to match the key point set and the sub-feature set, and the vector distance calculation with quadratic complexity is performed. The matching pairs whose minimum matching distance is less than 0.9 times the second minimum matching distance are retained, and the matching set is obtained.
[0053] S5 calculates the matched outer points and inner points according to the matching set, and performs identity authentication according to the number of inner points, thereby achieving fundus image matching.
[0054] According to the matching set, the RANSAC algorithm or the median method is used to calculate the matching outer and inner points, and the number of remaining inner points is used as the identity authentication indicator. If the indicator is less than the preset threshold, the identity authentication is rejected, otherwise the identity authentication is accepted; or the RANSAC algorithm or the median method is directly solved to obtain the affine matrix that describes the changes in the fundus image, and the image registration is obtained through the affine matrix. The matrix multiplication between the affine matrix and the coordinate point can transform the coordinate point to the coordinate system under the new perspective. The affine transformation represented by this affine matrix is performed on each pixel of the image to obtain the registration map.
[0055] In this embodiment, a PKE technology is proposed for training a key point detection model, which gradually expands key points with high repeatability and high reliability in a semi-supervised manner, solves the problem of insufficient manually labeled data, and effectively enhances the performance of image matching; a deep fundus image model is designed, which is more suitable for fundus image matching and has better performance than previous feature descriptors.
[0056] Embodiment 2
[0057] In order to verify the effectiveness of the solution in Example 1, this example conducts a large number of experiments on the two sub-problems of registration and identity recognition. The registration task is tested on a public dataset FIRE (FIRE is a retinal fundus image dataset, containing 129 fundus retinal images, which are combined into 134 pairs of image combinations by different features. These image combinations are divided into 3 categories according to their characteristics. The fundus images are collected by a Nidek AFC-210 fundus camera with a resolution of 2912x2912 and a visual elevation angle of 40 degrees. The images are jointly constructed by Papageorgiou Hospital and Aristotle University of Thessaloniki), and the test scheme follows the test scheme of GLAMpoints (Greedily Learned Accurate Matchpoints) and the test scheme of FIRE. The former is to calculate the registration failure rate, misalignment rate and acceptance rate based on the median and maximum values of the distance error between the query key point and the associated key point, and the latter is to make a graph of acceptance rate versus threshold based on the average value of the distance error, and calculate the normalized area under the X-axis (AUC) of the graph of acceptance rate versus threshold. The larger the value, the better. The X-axis is the manually set error threshold (pixel). For example, when X=5 and Y=0.8, it means that when the error threshold of the average distance is set to 5 pixels, the registration success rate is 80%, that is, 80% of the key point distances of the image pairs to be tested have an average value less than 5 pixels.
[0058] The identification task uses the most commonly used equal error rate (EER) indicator, that is, the point where the false acceptance rate and the false rejection rate are equal. Tests were conducted on three datasets: VARIA, CLINICAL, and BES. VARIA is a fundus image that only contains the optic disc area, and the fundus image is clear. CLINICAL is an internal dataset that mainly consists of fundus images with fundus diseases. BES is collected across multiple years and is closest to the real scene.
[0059] Table 1 is the ablation experiment data. The first row in the table uses PKE technology, while the second row does not use PKE technology; the third row is based on the first row, and the description decoder of the model in the first row is replaced with RootSIFT, and the fourth row is replaced with SOSNet. The performance comparison between the first row and the second row shows that the use of PKE technology can greatly improve the performance of the present invention in registration tasks and recognition tasks, verifying the effectiveness of PKE technology; the first row and the third and fourth rows mainly compare the performance between the description decoder of the present invention and other mainstream description decoders, which can verify the effectiveness of the network structure designed by the present invention and the training strategy of the description decoder, which not only reduces the amount of training calculations, but also does not lose the matching accuracy.
[0060] Table 1 Ablation experiment table of PKE technology and feature description decoder
[0061]
[0062] The performance comparison with other image matching methods is shown in Table 2. The experimental results show that the various indicators of the method in Example 1 are better than other matching schemes, and show obvious advantages on various data sets, while achieving high performance in both the registration task and the identity recognition task.
[0063] Table 2 Comparison table of the method in Example 1 and other matching methods
[0064]
[0065] Embodiment 3
[0066] Based on the same inventive concept, this embodiment discloses a fundus image matching system based on deep learning, including:
[0067] A model training module, used for inputting the fundus image into a deep fundus image model for training to obtain a trained deep fundus image model, wherein the deep fundus image model includes an encoder network and two decoder networks;
[0068] The model output module is used to input the fundus image to be tested into the trained deep fundus image model to generate a key point probability map and a description feature map;
[0069] A feature set extraction module is used to obtain a key point set and a sub-feature set according to a key point probability map and description features;
[0070] A matching set generation module is used to match the key point set and the sub-feature set to obtain a matching set;
[0071] The identity authentication module is used to calculate the matched outer points and inner points according to the matching set, and to perform identity authentication according to the number of inner points, so as to achieve fundus image matching.
[0072] Embodiment 4
[0073] Based on the same inventive concept, this embodiment discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform a fundus image matching method based on any of the above-mentioned deep learning methods.
[0074] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0075] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0076] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0077] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not deviate from the spirit and scope of the present invention should be included in the protection scope of the claims of the present invention. The above content is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.
Claims
1. A fundus image matching method based on deep learning, characterized in that: include: Inputting the fundus image into a deep neural network model for training to obtain a deep fundus image model that simultaneously realizes key point detection and feature extraction, wherein the deep fundus image model includes an encoder network and two decoder networks; The encoder network is responsible for extracting fundus image features, and the two decoder networks include a key point detection decoder and a feature description decoder, and the key point detection decoder is responsible for completing image key point detection; The feature description decoder is responsible for completing the key point feature description; The key point detection decoder uses progressive key point amplification technology to extract key points, and the extracted key points do not include key point annotation information; Key point extraction is screened using a two-stage filtering method: the first stage is geometry-based screening, which requires that key points that appear in the fundus image should also appear in the transformation map; The second stage is content-based screening, which requires that the key point features detected in the fundus image and the key point features in the transformation graph meet the optimal KNN matching, and the key points screened twice are added to the candidate key point set of Yt; Input the fundus image to be tested into a deep neural network model that simultaneously realizes key point detection and feature extraction, and generates a key point probability map and a description feature map; The key point probability map is generated by upsampling the high-level features in the key point detection decoder to obtain the same height and width as the low-level features in the key point detection decoder, splicing the high-level features and the low-level features along the channel dimension, and then convolving and upsampling the spliced features to obtain an output map with the same size as the input image, and generating a key point probability map after sigmoid activation; the description feature map is directly obtained by a quadratic interpolation method, and its size is consistent with the original image; Obtain a key point coordinate set and a key point feature set according to the key point probability map and the description feature map; Matching the key point coordinate set and the key point feature set corresponding to the two fundus images to be matched to obtain a matching set; Parameters of affine transformation are calculated according to the matching set to achieve image registration, matched external points and internal points are determined according to the matching set, and identity authentication is performed according to the number of the internal points.
2. The fundus image matching method according to claim 1, characterized in that: The fundus image only needs to have key point annotation information on a part of it, and the fundus image is input into the deep fundus image model. If the fundus image has key point annotation information, the key point annotation information is used as Input to the keypoint detection decoder.
3. The fundus image matching method according to claim 1 or 2, characterized in that: The feature description decoder obtains a key point set based on the key point detection decoder. Sampling is performed on the description feature map to obtain the feature set, and then based on the random affine transformation matrix Construct a triplet loss function for training.
4. The fundus image matching method according to claim 1 or 2, characterized in that: The matching set uses the RANSAC algorithm or the median method to calculate the matching external points and internal points, and the number of remaining internal points is used as an indicator of identity authentication. If the indicator is less than a preset threshold, the identity authentication is rejected, otherwise the identity authentication is accepted; or the RANSAC algorithm or the median method is directly solved to obtain an affine matrix that describes the changes in the fundus image, and image alignment is obtained through the affine matrix.
5. A fundus image matching system based on deep learning, used to implement the fundus image matching method according to any one of claims 1 to 4, characterized in that: include: A model training module, used for inputting the fundus image into a deep neural network model for training to obtain a deep fundus image model that simultaneously realizes key point detection and feature extraction, wherein the deep fundus image model includes an encoder network and two decoder networks; A model output module is used to input the fundus image to be tested into a deep neural network model that simultaneously realizes key point detection and feature extraction, and generates a key point probability map and a description feature map; A feature set extraction module is used to obtain a key point coordinate set and a key point feature set according to a key point probability map and description features; A matching set generation module, used for matching the key point coordinate set and the key point feature set corresponding to the two fundus images to be matched to obtain a matching set; The identity authentication module is used to calculate the parameters of the affine transformation according to the matching set to realize image registration, determine the matched external points and internal points according to the matching set, and perform identity authentication according to the number of the internal points.
6. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform the deep learning fundus image matching method according to any one of claims 1 to 4.