Dark light matching image feature matching method, system and equipment

Abstract

The invention relates to the field of computer vision, and discloses a dark light matching image feature matching method, system and device, and the method comprises the steps: S1, obtaining an original dark light matching image; s2, performing feature enhancement on the obtained original dark light matching image through an MSRCR feature enhancement algorithm to obtain an enhanced image; s3, extracting feature points of the enhanced image through a SuperPoint feature extraction model, and generating corresponding descriptors; meanwhile, the positions of feature points are recognized, descriptors are processed, and then the feature points are input into a SuperGlue neural network model; and S4, the SuperGlue neural network model carries out feature matching based on the input feature point positions and the processed descriptors and filters out mismatching points. According to the method, the problems that a traditional matching algorithm and a SuperGlue matching algorithm are small in matching number and low in precision in the dark light image are solved, the robustness is high, and the rotation pose and the translation pose are closer to true values.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a dark-light matching image feature matching method, system and device. Background Art

[0002] In visual odometry, feature matching, as the middle end connecting the data acquisition front end and the industrial production back end, plays an important role in accurately estimating the camera pose and restoring the real 3D information. Traditional matching methods such as Scale Invariant Feature Transform (SIFT), Oriented FAST and Rotated BRIEF (ORB), and Speed ​​Up Robust Features (SURF) are usually used to extract, match, and estimate the correspondence between feature points and the camera pose. In 2020, Paul-Edouard Sarlin et al. proposed the SuperGlue matching algorithm, which uses an end-to-end learning method. The algorithm combines the feature extraction algorithm SuperPoint, convolutional neural network, and attention mechanism, and uses the classic Sinkhorn algorithm to iteratively solve the matching problem, thereby improving the matching accuracy and robustness. However, both traditional feature matching algorithms and SuperGlue neural networks based on deep learning are often only studied for normal lighting. In dark light matching image matching, there are situations where the matching accuracy is low or even impossible to match, which leads to low accuracy or even impossible to estimate the camera pose.

[0003] In order to solve this problem, this patent proposes a dark light matching image feature matching method. First, the original low-light image pair is enhanced by MSRCR features, and the enhanced low-light image is extracted with the SuperPoint algorithm to generate descriptors, and then the SuperGlue matching algorithm is used to perform feature matching and filter out mismatched points. The difficulty of extracting feature points in dark light matching images is successfully solved, the extraction efficiency and accuracy of feature extraction operators are improved, thereby improving the matching accuracy, the number of extracted feature points and the number of correctly matched feature points are significantly increased, and the rotation and translation errors are greatly reduced. Summary of the invention

[0004] In view of the above-mentioned technical deficiencies, the technical problem to be solved by the present invention is to provide a dark-light matching image feature matching method, which aims to solve the problem that the existing traditional feature matching algorithm and the SuperGlue neural network model based on deep learning have low matching accuracy or even cannot match in dark-light matching image matching, thereby causing the problem of low accuracy or even cannot estimate the camera pose estimation.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a dark light matching image feature matching method, comprising:

[0006] S1, obtaining the original dark light matching image;

[0007] S2, performing feature enhancement on the acquired original dark light matching image by using the MSRCR feature enhancement algorithm to obtain an enhanced image;

[0008] S3, extracting feature points of the enhanced image through the SuperPoint feature extraction model and generating corresponding descriptors; at the same time, identifying the position of the feature points and processing the descriptors and inputting them into the SuperGlue neural network model;

[0009] S4, SuperGlue neural network model performs feature matching based on the input feature point positions and processed descriptors and filters out mismatched points.

[0010] Furthermore, the step S2 includes:

[0011] The MSRCR feature enhancement algorithm decomposes the original image into illumination component and reflection component, and introduces Gaussian filtering to filter out the illumination component. The enhancement formula is as follows:

[0012]

[0013] Where N is the number of scales, ω k is the weight of the corresponding scale, G k (x, y) is the Gaussian filter function corresponding to the kth channel, S i (x, y) is the original dark light matching image information, C i (x,y) is the color restoration factor of the i-th channel.

[0014] Furthermore, the step S3 includes:

[0015] The enhanced image is input into the encoder and the reduced-dimensional feature map is output to the decoder through spatial pooling.

[0016] The decoder is based on the reduced-dimensional feature map output by the encoder, and introduces the feature point detection head and the descriptor head to calculate the tensor dimensions respectively. And output the final tensor dimension R H×W , to identify the location of the feature point and the descriptor of the feature point at that location; where H and W are the height and width of the output tensor, 65 is the dimension of the probability distribution of each location, and 65 is a size of 8×8, and the last bit is discarded in the softmax process;

[0017] The semi-dense descriptor of the feature point at this position is bicubic interpolated, then L2 regularized and output as a dense map of the L2 normalized fixed-length descriptor.

[0018] Further, the step S4 includes:

[0019] Use the multi-layer perceptron MLP to aggregate the feature point positions and the descriptors after dense mapping. The specific formula is as follows:

[0020]

[0021] in, represents the feature point of the i-th aggregated visual descriptor and position information, d i represents the visual descriptor, P i Indicates location information;

[0022] Input the feature points of the i-th aggregated visual descriptor and position information into the attention graph neural network GNN;

[0023] The attention graph neural network GNN generates feature descriptors by iterative aggregation within and between graphs through self-attention and cross-attention and inputs them into the optimal matching layer;

[0024] The optimal matching layer calculates the inner product score based on the feature descriptors aggregated by the attention graph neural network GNN to perform feature matching.

[0025] Furthermore, the attention graph neural network GNN starts with self-attention to perform intra-graph aggregation and then uses cross-attention to perform inter-graph aggregation, thereby repeatedly iterating and looping to achieve a high degree of information aggregation; wherein the updated residual information in the graph is transmitted according to the following formula:

[0026]

[0027] α ij =softmax j (q i Τ k i )

[0028] Among them, m ε→i L is the result after aggregating all feature points. M is the intermediate expression of the iNth element in the image A at the lOth level, α ij P is the feature similarity, the larger the value, the more similar it is. j Q is attribute retrieval;

[0029] After L iterations, the feature descriptor is generated and input into the optimal matching layer. The specific formula is as follows:

[0030]

[0031] in, is the A feature descriptor, is the B feature descriptor.

[0032] Furthermore, the feature descriptor obtained by aggregation using the attention graph neural network GNN and The formula for calculating the inner product score is as follows:

[0033] S i,j = <f i A ,f i B >

[0034] Construct a score matrix based on the calculated inner product scores.

[0035] Furthermore, it also includes introducing a trash bin in the score matrix to collect unmatched feature points and incorrectly matched feature points, placing the unmatched feature points and incorrectly matched feature points in a separate set, and using the Sinkhorn algorithm to solve the optimal transmission problem to obtain the final matching result.

[0036] A system for a dark light matching image feature matching method, comprising:

[0037] Acquisition module, obtaining the original dark light matching image;

[0038] The enhancement module performs feature enhancement on the acquired original dark light matching image through the MSRCR feature enhancement algorithm to obtain an enhanced image;

[0039] The extraction module extracts the feature points of the enhanced image through the SuperPoint feature extraction model and generates the corresponding descriptors; at the same time, it identifies the location of the feature points and processes the descriptors before inputting them into the SuperGlue neural network model;

[0040] In the matching module, the SuperGlue neural network model performs feature matching based on the input feature point positions and processed descriptors and filters out mismatched points.

[0041] A computing device comprises: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions, which, when executed by the processor, implement the steps of a dark light matching image feature matching method.

[0042] The beneficial effects of the present invention are:

[0043] 1. The camera pose estimation accuracy of the present invention is significantly improved compared with the SuperGlue neural network model, and the robustness is improved;

[0044] 2. The present invention has high matching accuracy;

[0045] 3. The present invention solves the problem of small number of matches and low precision in dark light images of traditional matching algorithms and SuperGlue matching algorithms, improves robustness, and makes the rotational pose and translational pose closer to the true value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 work.

[0047] Figure 1 A flowchart of a dark light matching image feature matching method provided in Example 1 of the present invention.

[0048] Figure 2 This is a matching comparison diagram of Experiment 1 in Example 2 of the present invention.

[0049] Figure 3 This is a matching comparison diagram of Experiment 2 in Example 2 of the present invention.

[0050] Figure 4 This is a matching comparison diagram of Experiment 3 in Example 2 of the present invention.

[0051] Figure 5 This is a matching comparison diagram of Experiment 4 in Example 2 of the present invention.

[0052] Figure 6 This is a matching comparison diagram of Experiment 5 in Example 2 of the present invention.

[0053] Figure 7 This is a matching comparison diagram of Experiment 6 in Example 2 of the present invention.

[0054] Figure 8 This is a score curve diagram of the number of feature point matches and the matching accuracy in Example 2 of the present invention. DETAILED DESCRIPTION

[0055] 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.

[0056] Example 1

[0057] Reference Figure 1 This embodiment provides a dark light matching image feature matching method, comprising the following steps:

[0058] S1. Obtain the original dark-light matching image.

[0059] S2. The original dark light matching image is enhanced by the MSRCR feature enhancement algorithm to obtain an enhanced image. It should be noted that:

[0060] The MSRCR feature enhancement algorithm decomposes the original dark light matching image into illumination component and reflection component, and introduces Gaussian filtering to filter out the illumination component; the SSR algorithm decomposes the image into illumination component and reflection component. The core is to suppress or eliminate the low-frequency component of the image by some method, use the reflection image to reflect the image information, and then continue to white balance adjustment to enhance the image to obtain useful information of the image. Therefore, the Gaussian function is introduced to filter the input image. The Gaussian function is a good normalized low-pass filter. The Gaussian filter can filter out the illumination component. The working principle is as follows:

[0061] S(x,y)=L(x,y)·R(x,y)

[0062] Among them, S(x, y) is the original dark light matching image information, L(x, y) is the illumination component, and R(x, y) is the reflection component. In order to separate the illumination component and the reflection component and simplify the calculation, the logarithm operation is performed on both sides of the formula to obtain the following formula:

[0063] log(S(x,y))=log(L(x,y))+log(R(x,y))

[0064] Gaussian filtering is introduced to solve the lighting component L(x,y). The filtering calculation is as follows:

[0065] L(x,y)=S(x,y)·G(x,y)

[0066]

[0067] ∫∫G(x,y)dxdy=1

[0068] Where G(x,y) is a Gaussian function, x and y are the horizontal and vertical offsets relative to the center point, and σ is the standard deviation;

[0069] The MSR algorithm is equivalent to the weighted average of multiple SSR algorithms. Each filter channel is assigned a specific weight, and finally weighted fusion is performed to achieve feature enhancement. The principle is:

[0070]

[0071] Where N is the number of scales, ω k is the weight of the corresponding scale, S i (x, y) is the original dark light matching image information and satisfies the following formula:

[0072]

[0073] G k (x, y) is the Gaussian filter function corresponding to the kth channel, and its expression is:

[0074]

[0075] Among them, σ k is the Gaussian standard deviation on the kth scale, also known as the scale parameter. The values ​​of each Gaussian function channel can be adjusted as needed to achieve the experimental effect. Usually, the three channels R, G, and B are selected.

[0076] Based on the MSR algorithm, a color restoration factor is introduced and a multi-scale adaptive gain MSRCR is proposed to solve the color distortion problem. The principle is:

[0077]

[0078] Among them, C i (x,y) is the color restoration factor of the i-th channel, and the expression is as follows:

[0079] C i (x,y)=β(log(αS i (x,y))-log(Σ i∈(r,g,b) S i (x,y)))

[0080] Among them, α and β are gain coefficients, r, g, b are the three channels of R, G, and B.

[0081] S3, extract the feature points of the enhanced image through the SuperPoint feature extraction model and generate the corresponding descriptor; at the same time, identify the location of the feature points and process the descriptor before inputting it into the SuperGlue neural network model. It should be noted that:

[0082] The enhanced image is input into the encoder and the reduced-dimensional feature map is output to the decoder through spatial pooling.

[0083] Specifically, the encoder is a simple VGG model. The enhanced image H*W*1 is input into the encoder, and after passing through the convolution layer, pooling layer and non-linear activation function, the encoder outputs a tensor of H / 8*W / 8*128. The encoder aims to achieve dimensionality reduction through spatial pooling.

[0084] The decoder is based on the reduced-dimensional feature map output by the encoder. After the encoder reduces the dimension, it outputs a tensor of H / 8*W / 8*128. At this time, a convolution operation is required to upsample it back to full resolution. However, the upconvolution operation will increase the amount of calculation, so the feature point detection head and the descriptor head are designed to calculate the tensor dimensions separately. And output tensor dimension R H×W , tensor dimension R H×W The probability of each position being identified as a feature point in is used to identify the feature point position and the descriptor of the feature point at that position; where 65 is the size of 8×8, and the last bit is discarded in the softmax process;

[0085] By using a model such as UCN, a semi-dense descriptor of the feature point at the location is first taken as the network input, then bicubic interpolation is performed, followed by L2 regularization and output of a dense map of L2 normalized fixed-length descriptors.

[0086] S4, SuperGlue neural network model performs feature matching based on the input feature point positions and processed descriptors and filters out mismatched points. It should be noted that:

[0087] Use the multi-layer perceptron MLP to aggregate the feature point positions and the descriptors after dense mapping. The specific formula is as follows:

[0088]

[0089] Among them, (0) Xi represents the feature point of the i-th aggregated visual descriptor and position information, d i represents the visual descriptor, P i Indicates location information;

[0090] Input the feature points of the i-th aggregated visual descriptor and position information into the attention graph neural network GNN;

[0091] The attention graph neural network GNN generates feature descriptors through iterative aggregation within and between graphs through self-attention and cross-attention and inputs them into the optimal matching layer;

[0092] The best matching layer calculates the inner product score based on the feature descriptors aggregated by the attention graph neural network GNN for feature matching.

[0093] Specifically, the attention graph neural network GNN starts with self-attention for intra-graph aggregation and then uses cross-attention for inter-graph aggregation, repeatedly iterating to achieve a high degree of information aggregation; wherein the updated residual information in the graph is transmitted according to the following formula:

[0094]

[0095] α ij =softmax j (q i Τ k i )

[0096] Among them, m ε→i L is the result after aggregating all feature points; M is the intermediate expression of the iNth element in the image A at the lOth level; α ij P is the feature similarity, the larger the value, the more similar it is; V j Q is attribute retrieval;

[0097] After L iterations, the feature descriptor is generated and input into the optimal matching layer. The specific formula is as follows:

[0098]

[0099] in, is the A feature descriptor, is the B feature descriptor.

[0100] Furthermore, the best matching layer is the second main module of the SuperGlue neural network model. The feature descriptors obtained by the attention graph neural network GNN are input into the best matching layer network, and then the shortest path calculation score matrix is ​​calculated according to the shortest path problem and 0 and 1 are assigned to achieve feature matching. The best matching layer network calculates a score matrix To represent some potential matches, but because the workload of establishing separate representations for M×N potential matches is huge, this embodiment uses the feature descriptor obtained by aggregation of the attention graph neural network GNN and Calculate the inner product score, the formula is as follows:

[0101] S i,j = <f i A ,f i B >

[0102] In order to solve the occlusion and visibility problems, a garbage bin is introduced into the score matrix to collect unmatched feature points and incorrectly matched feature points. The unmatched feature points and incorrectly matched feature points are placed in a separate set, and the Sinkhorn algorithm is used to solve the optimal transmission problem to obtain the final matching result.

[0103] This embodiment also provides a dark light matching image feature matching system, including:

[0104] Acquisition module, obtaining the original dark light matching image;

[0105] The enhancement module performs feature enhancement on the acquired original dark light matching image through the MSRCR feature enhancement algorithm to obtain an enhanced image;

[0106] The extraction module extracts the feature points of the enhanced image through the SuperPoint feature extraction model and generates the corresponding descriptors; at the same time, it identifies the location of the feature points and processes the descriptors before inputting them into the SuperGlue neural network model;

[0107] In the matching module, the SuperGlue neural network model performs feature matching based on the input feature point positions and processed descriptors and filters out mismatched points.

[0108] This embodiment further provides a computing device, applicable to a dark light matching image feature matching method, including:

[0109] A memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement a dark light matching image feature matching method as proposed in the above embodiment.

[0110] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0111] Example 2

[0112] Reference Figure 2 to Figure 8 This embodiment is the second embodiment of the present invention. This embodiment is different from the first embodiment in that it provides a verification test of a dark light matching image feature matching method to verify and illustrate the technical effects used in this method.

[0113] Specifically, the following steps are included:

[0114] A1. Obtain image matching data. We selected the ScanNet indoor dataset, which contains images of the same scene with different perspectives, ground truth poses and depths. To meet the experimental requirements, we also manually adjusted the brightness and exposure of the original low-light image to obtain the original dark-light matching image.

[0115] A2. Calculate the transformation relationship between the original dark light matching images through the SuperGlue neural network model, calculate the relative rotation and translation errors of the two original dark light matching images according to the transformation matrix, match the two original dark light matching images to obtain the matching result and calculate the original dark light matching score Mscore1;

[0116] The transformation matrix in step A2 is shown in formula (1):

[0117]

[0118] in, is the antisymmetric matrix of the translation matrix, and R is the rotation matrix.

[0119] The original dark-light matching score Mscore1 is shown in formula (2):

[0120]

[0121] Among them, num_correct is the number of correct matches, and len(pts11) is the total number of matches.

[0122] A3. The original dark light matching image is enhanced by the MSRCR feature enhancement algorithm. The original dark light matching image information is decomposed into illumination component and reflection component. The illumination component is filtered out by Gaussian function to enhance the image texture information. The brightness and contrast of the image are improved by white balance.

[0123] The enhancement formula in step A3 is shown in formula (3):

[0124]

[0125] Where N is the number of scales, ω k is the weight of the corresponding scale, G k (x, y) is the Gaussian filter function corresponding to the kth channel, Si (x, y) is the original dark light matching image information, C i (x,y) is the color restoration factor of the i-th channel.

[0126] A4. Extract feature points of the enhanced image through the SuperPoint feature extraction model and generate corresponding descriptors; at the same time, identify the location of the feature points and process the descriptors before inputting them into the SuperGlue neural network model;

[0127] A5. Calculate the transformation relationship between the images enhanced by S3 through the SuperGlue neural network model, determine the spatial relative positions between the same-name points, and calculate the enhanced matching score Mscore2;

[0128] A6, using the enhanced matching score Mscore2 obtained in A5, verifying the matching performance of the method on the enhanced matching score Mscore2, and analyzing the performance of the improved method using the number of matched feature points and the number of correctly matched feature points;

[0129] like Figure 2 to Figure 8 As shown, this embodiment experiments on 6 groups of matches, ΔR represents the rotation error, Δt represents the translation error, and the Inliers in the sample points contain the number of correctly matched feature points and the total number of matched feature points, and also provides the number of feature points extracted from the left and right images and the number of feature points involved in the matching.

[0130] In order to analyze the accuracy of the SuperGlue+MSRCR method, it is compared with the camera pose estimation accuracy of the original SuperGlue algorithm. The specific results are shown in Table 1.

[0131] Table 1: Matching accuracy table.

[0132]

[0133]

[0134] From Table 1, Figure 2 to Figure 8 It can be seen that:

[0135] (1) From the perspective of the number of matches, for dark-light matching images, the number of correctly matched and totally matched feature points of the traditional original SuperGlue neural network model is relatively small, while the number of correctly matched and totally matched feature points of the matching method using SuperGlue+MSRCR in the present invention is significantly increased. The number of correctly matched feature points of SuperGlue+MSRCR is approximately twice the number of correctly matched specific numbers of the original SuperGlue neural network model, and the number of totally matched feature points of SuperGlue+MSRCR is approximately 1.5 times the number of totally matched specific numbers of the original SuperGlue.

[0136] (2) After enhancing the original dark light matching image, the number of feature points extracted is significantly increased and the matching accuracy is significantly improved.

[0137] (3) For some poor quality and blurred dark light matching image pairs, after feature enhancement by the MSRCR feature enhancement algorithm, the texture information of the image is enhanced, and the brightness and contrast of the image are improved. The original SuperGlue neural network model cannot complete the matching task, but the matching method of SuperGlue+MSRCR used in the present invention can still achieve feature matching.

[0138] From the above conclusions, it can be seen that compared with traditional algorithms and deep learning algorithms, the method proposed in the present invention can achieve higher-precision camera pose estimation. After feature enhancement through the MSRCR feature algorithm, the texture information of the dark-light matching image is enhanced, and the brightness and contrast of the original dark-light matching image are improved, thereby improving the camera pose.

[0139] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A dark light matching image feature matching method, characterized in that: include: S1, obtaining the original dark light matching image; S2, performing feature enhancement on the acquired original dark light matching image by using the MSRCR feature enhancement algorithm to obtain an enhanced image; S3, extracting feature points of the enhanced image through the SuperPoint feature extraction model and generating corresponding descriptors; at the same time, identifying the position of the feature points and processing the descriptors and inputting them into the SuperGlue neural network model; S4, SuperGlue neural network model performs feature matching based on the input feature point positions and processed descriptors and filters out mismatched points.

2. A dark light matching image feature matching method as claimed in claim 1, characterized in that: The step S2 comprises: The MSRCR feature enhancement algorithm decomposes the original image into illumination component and reflection component, and introduces Gaussian filtering to filter out the illumination component. The enhancement formula is as follows: Where N is the number of scales, ω k is the weight of the corresponding scale, G k (x, y) is the Gaussian filter function corresponding to the kth channel, S i (x, y) is the original dark light matching image information, C i (x,y) is the color restoration factor of the i-th channel.

3. A dark light matching image feature matching method as claimed in claim 1, characterized in that: The step S3 comprises: The enhanced image is input into the encoder and the reduced-dimensional feature map is output to the decoder through spatial pooling. The decoder is based on the reduced-dimensional feature map output by the encoder, and introduces the feature point detection head and the descriptor head to calculate the tensor dimensions respectively. And output the final tensor dimension R H×W , to identify the location of the feature point and the descriptor of the feature point at that location; where H and W are the height and width of the output tensor, 65 is the dimension of the probability distribution of each location, and 65 is a size of 8×8, and the last bit is discarded in the softmax process; The semi-dense descriptor of the feature point at this position is bicubic interpolated, then L2 regularized and output as a dense map of the L2 normalized fixed-length descriptor.

4. The dark light matching image feature matching method according to claim 1, characterized in that: The step S4 comprises: Use the multi-layer perceptron MLP to aggregate the feature point positions and the descriptors after dense mapping. The specific formula is as follows: in, represents the feature point of the i-th aggregated visual descriptor and location information, d i represents the visual descriptor, P i Indicates location information; Input the feature points of the i-th aggregated visual descriptor and position information into the attention graph neural network GNN; The attention graph neural network GNN generates feature descriptors by iterative aggregation within and between graphs through self-attention and cross-attention and inputs them into the optimal matching layer; The optimal matching layer calculates the inner product score based on the feature descriptors aggregated by the attention graph neural network GNN to perform feature matching.

5. A dark light matching image feature matching method as claimed in claim 4, characterized in that: The attention graph neural network GNN starts with self-attention for intra-graph aggregation and then uses cross-attention for inter-graph aggregation, and iterates repeatedly to achieve high aggregation of information; wherein the updated residual information in the graph is transmitted according to the following formula: m ε→i =S j:(i,j)∈ε a ij V j Among them, m ε→i L is the result after aggregating all feature points; is the intermediate expression of the iNth element in the image A at the lOth element; α ij P is the feature similarity, the larger the value, the more similar it is; V j Q is attribute retrieval; After L iterations, the feature descriptor is generated and input into the optimal matching layer. The specific formula is as follows: in, is the A feature descriptor, is the B feature descriptor.

6. A dark light matching image feature matching method as claimed in claim 4, characterized in that: Feature descriptors aggregated using attention graph neural network GNN and The formula for calculating the inner product score is as follows: S i,j =<f i A ,f i B > Construct a score matrix based on the calculated inner product scores.

7. A dark light matching image feature matching method as claimed in claim 6, characterized in that: It also includes introducing a trash bin into the score matrix to collect unmatched feature points and feature points with incorrect matching, placing the unmatched feature points and feature points with incorrect matching in a separate set, and using the Sinkhorn algorithm to solve the optimal transmission problem to obtain a final matching result.

8. A system based on the dark light matching image feature matching method as claimed in claim 1, characterized in that: Acquisition module, obtaining the original dark light matching image; The enhancement module performs feature enhancement on the acquired original dark light matching image through the MSRCR feature enhancement algorithm to obtain an enhanced image; The extraction module extracts the feature points of the enhanced image through the SuperPoint feature extraction model and generates the corresponding descriptors; at the same time, it identifies the location of the feature points and processes the descriptors before inputting them into the SuperGlue neural network model; In the matching module, the SuperGlue neural network model performs feature matching based on the input feature point positions and processed descriptors and filters out mismatched points.

9. A computing device comprising: Memory and processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of a dark light matching image feature matching method described in any one of claims 1 to 7 are implemented.

Images