Translation and rotation invariant image feature coding method and system
By combining Fourier transform and polar coordinate transformation in image processing, the difficulty of feature matching caused by image viewing angle differences is solved, the translation and rotation invariance of the image are achieved, and the accuracy and efficiency of feature matching are improved.
Patent Information
- Application Number
- CN202411759201.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-03-28
AI Technical Summary
In machine vision and vision-based site recognition tasks, it is difficult to match features due to differences in view angles, and prior art is difficult to effectively handle translation and rotation changes.
By combining Fourier transform and polar coordinate transformation, the translation invariance and rotation invariance of the image are achieved, and multi-angle rotation feature extraction and maximum pooling operation are combined to improve the expression ability and robustness of the model.
It realizes robustness to the translation and rotation changes of images, improves the accuracy and efficiency of feature matching, and is suitable for visual location recognition and other image processing tasks.
Smart Images
Figure CN119722437A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine vision and vision-based location recognition, and in particular to a translation and rotation-invariant image feature encoding method and system. Background Art
[0002] Image feature extraction is one of the most important technologies in the field of robotics and machine vision applications. In particular, in visual place recognition (VPR) tasks, image features often need to be retrieved and matched. However, due to the different access times, there is often a large difference in perspective between the query image and the stored image, which makes image matching from different sources difficult and challenging. For example, in unmanned driving, drone navigation, and multi-robot collaborative tasks, the ability to accurately identify the same place directly affects the reliability and efficiency of the system.
[0003] Due to the change in the altitude of the drone, the scale of the photos taken by it may be inconsistent with the Bird's Eye View (BEV) of the ground lidar point cloud. Even the photos taken at the same location may have scale changes due to the different flying altitudes of the drone. The ground lidar point cloud BEV image is obtained from the ground perspective, while the drone aerial photos are taken from a high altitude. The difference in viewpoints between the two may cause the shape, size and layout of objects in the image to be significantly different, which will increase the difficulty of matching. Because the resolution of photos taken by drones depends on the altitude of the drone and the quality of the camera, the difference in resolution will also affect the matching accuracy. Due to the existence of the above factors, it may be difficult to find reliable correspondences directly using traditional feature point matching methods (such as SIFT, SURF, etc.).
[0004] Although current deep learning methods have made significant progress in image feature extraction, they are still insufficient in processing images with large differences in perspective and completing feature matching well. In particular, there is still room for improvement in the representation of translation invariance and rotation invariance. Ground-based lidar point cloud BEV images are acquired from the perspective of the ground, while drone aerial photos are taken from a high altitude, which inevitably involve rotation and translation problems. Summary of the invention
[0005] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a translation and rotation invariant image feature encoding method and system. The translation and rotation invariant image feature encoding method and system provided by the present invention achieve strong translation invariance and rotation invariance by combining Fourier transform and polar coordinate transform. Through multi-angle rotation feature extraction and maximum pooling operations, information of different scales is indirectly captured, and the expression ability and robustness of the model are improved. The present invention is highly flexible and can flexibly select different backbone networks. It is not only advanced in visual place recognition tasks, but also has broad application prospects and important academic research value in other tasks such as image classification and semantic segmentation.
[0006] The present invention provides a translation and rotation invariant image feature encoding method, comprising: S1: Acquire an aerial image and a ground point cloud BEV image, and perform affine transformation on the aerial image and the ground point cloud BEV image to obtain an affine image; S2: Obtain a translation invariant image by performing Fourier transform on the affine image; By performing polar coordinate transformation on the affine image, a rotation invariant image is obtained; S3: Use convolutional neural networks to extract image features of affine images, image features of translation-invariant images, and image features of rotation-invariant images; S4: performing inverse Fourier transform on the image features of the translation invariant image to obtain inverse translation invariant image features; performing inverse polar coordinate transform on the image features of the rotation invariant image to obtain inverse rotation invariant image features; S5: performing an inverse affine rotation transformation on the image feature of the affine image to obtain a first image feature, performing an inverse affine rotation transformation on the inverse translation invariant image feature to obtain a second image feature, and performing an inverse affine rotation transformation on the inverse rotation invariant image feature to obtain a third image feature; S6: Repeat steps S2 to S5 times, obtained Group image features, fuse all image features to obtain fused features; S7: Obtain an attention fusion feature by performing attention calculation on the fusion feature; obtain a rotation invariant feature by upsampling and normalizing the attention fusion feature.
[0007] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further comprises: Affine transformation, we get A group of affine images, the calculation expression of the affine image is: in, is an affine image, is the random rotation angle, For aerial images and ground point cloud BEV images.
[0008] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method also includes step S2, wherein an affine image is input into three channels, channel A performs Fourier transform on the affine image to obtain a translation invariant image; channel B performs polar coordinate transform on the affine image to obtain a rotation invariant image; and channel C does not process the affine image.
[0009] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further includes: in step S3, using a CNN Backbone feature extraction method and a ResNet50 convolutional neural network to extract image features of affine images, image features of translation invariant images, and image features of rotation invariant images; The calculation expression of the image feature of the affine image is: in, is the image feature of the affine image, is the ResNet50 convolutional neural network function, Affine images, The calculation expression of the image features of the translation invariant image is: in, is the image feature of the translation invariant image, is a translation invariant image, The calculation expression of the image feature of the rotation invariant image is: in, is the image feature of the rotation invariant image, is a rotation-invariant image.
[0010] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further includes step S5: The calculation expression of the first image feature is: in, is the first image feature, is the random rotation angle, is the image feature of the affine image; The calculation expression of the second image feature is: in, is the second image feature, is the inverse translation invariant image feature; The calculation expression of the third image feature is: in, is the third image feature, It is the inverse rotation invariant image feature.
[0011] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further includes step S6, wherein the calculation expression of the fusion feature is: in, To fusion features, is the maximum pooling function, is the first image feature obtained by the first transformation, is the first image feature obtained by the second transformation, For the The first image feature obtained by the transformation is is the second image feature obtained by the first transformation, is the second image feature obtained by the second transformation, For the The second image feature obtained by the transformation is is the third image feature obtained by the first transformation, is the third image feature obtained by the second transformation, For the The third image feature obtained by the second transformation.
[0012] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further includes step S7, extracting features from fusion features to obtain a query matrix, a key matrix and a value matrix, obtaining attention fusion features through the query matrix, the key matrix and the value matrix, and calculating the attention fusion features as follows: in, is the attention fusion feature, is the query matrix, is the key matrix, is the dimension of the key vector, is the value matrix, is transposed.
[0013] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further includes generating a grid by expanding a unit affine transformation matrix, sampling the attention fusion feature using the generated grid, and generating a new image feature. The calculation expression of the new image feature is: in, is the new image feature, is the bicubic interpolation function, is an affine grid, It is the attention fusion feature.
[0014] According to a translation and rotation invariant image feature encoding method provided by the present invention, the method further includes obtaining a rotation invariant feature by performing L2 normalization on a new image feature; in, is a rotation invariant feature, is the number of channels, The horizontal axis of the feature, is the vertical coordinate of the feature, For the The channel coordinates are New image features.
[0015] The present invention also provides a translation and rotation invariant image feature encoding system, which is used to execute any of the above translation and rotation invariant image feature encoding methods, comprising: An affine transformation module, wherein the affine transformation module acquires an aerial image and a ground point cloud BEV image, and performs an affine transformation on the aerial image and the ground point cloud BEV image to obtain an affine image; A transformation module, wherein the transformation module obtains a translation-invariant image by performing Fourier transformation on the affine image; and obtains a rotation-invariant image by performing polar coordinate transformation on the affine image; A feature extraction module, wherein the feature extraction module uses a convolutional neural network to extract image features of an affine image, image features of a translation-invariant image, and image features of a rotation-invariant image; An inverse transformation module, wherein the inverse transformation module performs an inverse Fourier transformation on the image features of the translation invariant image to obtain an inverse translation invariant image feature; and performs an inverse polar coordinate transformation on the image features of the rotation invariant image to obtain an inverse rotation invariant image feature; An inverse affine module, wherein the inverse affine module obtains a first image feature by performing an inverse affine rotation transformation on the image feature of the affine image, obtains a second image feature by performing an inverse affine rotation transformation on the inverse translation invariant image feature, and obtains a third image feature by performing an inverse affine rotation transformation on the inverse rotation invariant image feature; Feature fusion module, the feature fusion module Perform feature fusion on group image features to obtain fusion features; The acquisition module obtains attention fusion features by performing attention calculation on the fusion features; and obtains rotation invariant features by upsampling and normalizing the attention fusion features.
[0016] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: The present invention provides a translation and rotation invariant image feature encoding method and system, which realizes strong translation invariance and rotation invariance by combining Fourier transform and polar coordinate transform. Through multi-angle rotation feature extraction and maximum pooling operation, information of different scales is indirectly captured, and the expression ability and robustness of the model are improved. The present invention is highly flexible and can flexibly select different backbone networks. It is not only advanced in visual place recognition tasks, but also has broad application prospects and important academic research value in other tasks such as image classification and semantic segmentation.
[0017] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] In order to more clearly illustrate the technical solutions in the present invention or 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 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.
[0019] Figure 1 It is a flow chart of a translation and rotation invariant image feature encoding method provided by the present invention.
[0020] Figure 2 It is a schematic diagram of a model framework of a translation and rotation invariant image feature encoding method provided by the present invention.
[0021] Figure 3 It is a structural schematic diagram of a translation and rotation invariant image feature encoding system provided by the present invention.
[0022] Reference numerals: 101. Affine transformation module; 102. Transformation module; 103. Feature extraction module; 104. Inverse transformation module; 105. Inverse affine module; 106. Feature fusion module; 107. Acquisition module. DETAILED DESCRIPTION
[0023] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme in the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.
[0024] In the description of the embodiments of the present invention, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples without contradiction.
[0025] Combine the following Figures 1 to 3 The present invention describes a translation and rotation invariant image feature encoding method and system.
[0026] like Figure 1 As shown, a translation and rotation invariant image feature encoding method comprises: S1: Acquire an aerial image and a ground point cloud BEV image, and perform affine transformation on the aerial image and the ground point cloud BEV image to obtain an affine image; Aerial images are obtained by taking photos taken by drones; ground point cloud BEV images are obtained by ground laser radar; Aerial images and ground point cloud BEV images Affine transformation, we get A group of affine images, the calculation expression of the affine image is: in, is an affine image, is the random rotation angle, For aerial images and ground point cloud BEV images.
[0027] S2: Obtain a translation invariant image by performing Fourier transform on the affine image; obtain a rotation invariant image by performing polar coordinate transform on the affine image; The affine image is input into three channels, and channel A performs Fourier transform on the affine image to obtain a translation invariant image; the calculation expression is: in, is a translation invariant image, is the width of the image, is the height of the image, is the horizontal coordinate of the image, is the vertical coordinate of the image, is an imaginary unit, The image coordinates in the frequency domain coordinate system, is the channel A affine image; The image is converted from the spatial domain to the frequency domain through Fourier transform. The translation invariance property of Fourier transform is utilized to enhance the robustness of the model to translation changes. Fourier transform can convert the translation changes in the image into phase changes, thereby maintaining the consistency of features in the frequency domain.
[0028] Channel B for affine image Perform polar coordinate transformation to obtain a rotation-invariant image; The calculation expression is: in, is a rotation invariant image, is the polar diameter, is the polar angle; By rotating the input image at multiple angles, the features at different rotation angles are extracted, and the features are extracted after conversion to the polar coordinate system. This multi-angle rotation feature and polar coordinate extraction method can capture image information at different rotation angles and improve the model's robustness to rotation changes.
[0029] Channel C does not process the affine image.
[0030] S3: Use convolutional neural networks to extract image features of affine images, image features of translation-invariant images, and image features of rotation-invariant images; Adopting CNN Backbone feature extraction method, ResNet50 convolutional neural network is used to extract image features of affine images, image features of translation-invariant images, and image features of rotation-invariant images; The calculation expression of the image feature of the affine image is: in, is the image feature of the affine image, is the ResNet50 convolutional neural network function, Affine images, The calculation expression of the image features of the translation invariant image is: in, is the image feature of the translation invariant image, is a translation invariant image, The calculation expression of the image feature of the rotation invariant image is: in, is the image feature of the rotation invariant image, is a rotation invariant image, In some specific embodiments of the present invention, the first 4 layers of the ResNet50 convolutional neural network are used. The convolutional neural network can select different backbone networks, such as ResNet34, ResNet50, etc., and adjust according to task requirements and computing resources. This flexibility makes the model more adaptable and scalable in different application scenarios, and can be easily integrated into the existing deep learning framework to improve development efficiency and model performance.
[0031] S4: performing inverse Fourier transform on the image features of the translation invariant image to obtain inverse translation invariant image features; performing inverse polar coordinate transform on the image features of the rotation invariant image to obtain inverse rotation invariant image features; The characteristics Perform inverse Fourier transform and convert from frequency domain back to spatial domain. The calculation expression is: in, is the inverse translation invariant image feature, The characteristics Converting from polar coordinates back to Cartesian coordinates, we get , the calculation expression is: in, It is the inverse rotation invariant image feature.
[0032] S5: performing an inverse affine rotation transformation on the image feature of the affine image to obtain a first image feature, performing an inverse affine rotation transformation on the inverse translation invariant image feature to obtain a second image feature, and performing an inverse affine rotation transformation on the inverse rotation invariant image feature to obtain a third image feature; The calculation expression of the first image feature is: in, is the first image feature, is the random rotation angle, is the image feature of the affine image; The calculation expression of the second image feature is: in, is the second image feature, is the inverse translation invariant image feature; The calculation expression of the third image feature is: in, is the third image feature, It is the inverse rotation invariant image feature.
[0033] S6S6: Repeat steps S2 to S5 times, obtained Group image features, Perform feature fusion on group image features to obtain fusion features; The calculation expression of fusion features is: in, To fusion features, is the maximum pooling function, is the first image feature obtained by the first transformation, is the first image feature obtained by the second transformation, For the The first image feature obtained by the transformation is is the second image feature obtained by the first transformation, is the second image feature obtained by the second transformation, For the The second image feature obtained by the transformation is is the third image feature obtained by the first transformation, is the third image feature obtained by the second transformation, For the The third image feature obtained by the second transformation.
[0034] In some embodiments of the present invention, .
[0035] Multi-angle rotation feature extraction and maximum pooling operations indirectly capture information of different scales and combine features of different scales to improve the expressiveness and robustness of the model.
[0036] S7: Obtain an attention fusion feature by performing attention calculation on the fusion feature; obtain a rotation invariant feature by performing upsampling and normalization on the attention fusion feature; Will The transformation is divided into several local patch blocks. Each local patch block obtains a feature vector after linear transformation. The present invention organizes these feature vectors into a matrix , where each row represents the feature vector of a local patch block. Next, based on this matrix To calculate the self-attention, Transform the matrix by linear query Get the query matrix , linear key value transformation matrix Get the key matrix , linear value transformation matrix Get the value matrix ; The calculation expression is: Using the Query Matrix , the bond matrix , the dimension of the key vector , and the value matrix ,pass The function obtains the attention fusion feature, and the calculation expression of the attention fusion feature is: in, is the attention fusion feature, is transposed.
[0037] Generate a grid by expanding the unit affine transformation matrix, and use the generated grid to transform the feature map Sampling to generate new image features ; The input matrix size is , by upsampling, the output matrix size is ; Affine Mesh The matrix size is , each pixel Position in the input tensor, position in the output tensor The calculation expression is: We use bicubic interpolation to upsample the input matrix and obtain , the calculation expression is: in, is the new image feature, is the bicubic interpolation function, is an affine mesh; By performing L2 normalization on the new image features, rotation invariant features are obtained; in, is a rotation invariant feature, is the number of channels, The horizontal axis of the feature, is the vertical coordinate of the feature, For the The channel coordinates are New image features.
[0038] Model framework such as Figure 2 As shown, the present invention combines Fourier transform and polar coordinate transform to achieve strong translation invariance and rotation invariance. Through multi-angle rotation feature extraction and maximum pooling operation, information of different scales is indirectly captured, and the expression ability and robustness of the model are improved. The present invention is highly flexible and can flexibly select different backbone networks. It is not only advanced in visual place recognition tasks, but also has broad application prospects and important academic research value in other tasks such as image classification and semantic segmentation.
[0039] like Figure 3 As shown, a translation and rotation invariant image feature encoding system is used for the above-mentioned translation and rotation invariant image feature encoding method, comprising: The affine transformation module 101 acquires the aerial image and the ground point cloud BEV image, and performs affine transformation on the aerial image and the ground point cloud BEV image to obtain an affine image; The transformation module 102 obtains a translation invariant image by performing Fourier transformation on the affine image; obtains a rotation invariant image by performing polar coordinate transformation on the affine image; The feature extraction module 103 uses a convolutional neural network to extract image features of affine images, image features of translation-invariant images, and image features of rotation-invariant images; The inverse transformation module 104 performs an inverse Fourier transformation on the image features of the translation invariant image to obtain an inverse translation invariant image feature; performs an inverse polar coordinate transformation on the image features of the rotation invariant image to obtain an inverse rotation invariant image feature; The inverse affine module 105 obtains the first image feature by performing an inverse affine rotation transformation on the image feature of the affine image, obtains the second image feature by performing an inverse affine rotation transformation on the inverse translation invariant image feature, and obtains the third image feature by performing an inverse affine rotation transformation on the inverse rotation invariant image feature; Feature fusion module 106 obtains Group image features, fuse all image features to obtain fused features; The acquisition module 107 obtains the attention fusion feature by performing attention calculation on the fusion feature; and obtains the rotation invariant feature by upsampling and normalizing the attention fusion feature.
[0040] Through the collaborative work of the above modules, combined with Fourier transform and polar coordinate transformation, strong translation invariance and rotation invariance are achieved. Through multi-angle rotation feature extraction and maximum pooling operations, information of different scales is indirectly captured, improving the expressiveness and robustness of the model.
[0041] 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 it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A translation and rotation invariant image feature encoding method, characterized in that: include: S1: Acquire an aerial image and a ground point cloud BEV image, and perform affine transformation on the aerial image and the ground point cloud BEV image to obtain an affine image; S2: Obtain a translation invariant image by performing Fourier transform on the affine image; obtain a rotation invariant image by performing polar coordinate transform on the affine image; S3: Use convolutional neural networks to extract image features of affine images, image features of translation-invariant images, and image features of rotation-invariant images; S4: performing inverse Fourier transform on the image features of the translation invariant image to obtain inverse translation invariant image features; performing inverse polar coordinate transform on the image features of the rotation invariant image to obtain inverse rotation invariant image features; S5: performing an inverse affine rotation transformation on the image feature of the affine image to obtain a first image feature, performing an inverse affine rotation transformation on the inverse translation invariant image feature to obtain a second image feature, and performing an inverse affine rotation transformation on the inverse rotation invariant image feature to obtain a third image feature; S6: Repeat steps S2 to S5 times, obtained Group image features, Perform feature fusion on group image features to obtain fusion features; S7: Obtain an attention fusion feature by performing attention calculation on the fusion feature; obtain a rotation invariant feature by upsampling and normalizing the attention fusion feature.
2. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: Aerial images and ground point cloud BEV images Affine transformation, we get A group of affine images, the calculation expression of the affine image is: in, is an affine image, is the angle of random rotation, For aerial images and ground point cloud BEV images.
3. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: In step S2, the affine image is input into three channels, channel A performs Fourier transform on the affine image to obtain a translation invariant image; channel B performs polar coordinate transform on the affine image to obtain a rotation invariant image; channel C does not process the affine image.
4. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: In step S3, the CNN Backbone feature extraction method is used to extract the image features of affine images, the image features of translation-invariant images, and the image features of rotation-invariant images using the ResNet50 convolutional neural network. The calculation expression of the image feature of the affine image is: in, is the image feature of the affine image, is the ResNet50 convolutional neural network function, Affine images, The calculation expression of the image features of the translation invariant image is: in, is the image feature of the translation invariant image, is a translation invariant image, The calculation expression of the image feature of the rotation invariant image is: in, is the image feature of the rotation invariant image, is a rotation-invariant image.
5. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: In step S5, The calculation expression of the first image feature is: in, is the first image feature, is the random rotation angle, is the image feature of the affine image; The calculation expression of the second image feature is: in, is the second image feature, is the inverse translation invariant image feature; The calculation expression of the third image feature is: in, is the third image feature, It is the inverse rotation invariant image feature.
6. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: In step S6, the calculation expression of the fusion feature is: in, To fusion features, is the maximum pooling function, is the first image feature obtained by the first transformation, is the first image feature obtained by the second transformation, For the The first image feature obtained by the transformation is is the second image feature obtained by the first transformation, is the second image feature obtained by the second transformation, For the The second image feature obtained by the transformation is is the third image feature obtained by the first transformation, is the third image feature obtained by the second transformation, For the The third image feature obtained by the second transformation.
7. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: In step S7, feature extraction is performed on the fusion feature to obtain a query matrix, a key matrix, and a value matrix. The attention fusion feature is obtained through the query matrix, the key matrix, and the value matrix. The calculation expression of the attention fusion feature is: in, is the attention fusion feature, is the query matrix, is the key matrix, is the dimension of the key vector, is the value matrix, is transposed.
8. The method for encoding image features invariant to translation and rotation according to claim 1, characterized in that: By expanding the unit affine transformation matrix, a grid is generated, and the generated grid is used to sample the attention fusion features to generate new image features. The calculation expression of the new image features is: in, is the new image feature, is the bicubic interpolation function, is an affine grid, It is the attention fusion feature.
9. The method for encoding image features invariant to translation and rotation according to claim 8, characterized in that: By performing L2 normalization on the new image features, rotation invariant features are obtained; in, is a rotation invariant feature, is the number of channels, The horizontal axis of the feature, is the vertical coordinate of the feature, For the The channel coordinates are New image features.
10. A translation and rotation invariant image feature coding system, used to execute a translation and rotation invariant image feature coding method as claimed in any one of claims 1 to 9, characterized in that: include: An affine transformation module, wherein the affine transformation module acquires an aerial image and a ground point cloud BEV image, and performs an affine transformation on the aerial image and the ground point cloud BEV image to obtain an affine image; A transformation module, wherein the transformation module obtains a translation-invariant image by performing Fourier transformation on the affine image; and obtains a rotation-invariant image by performing polar coordinate transformation on the affine image; A feature extraction module, wherein the feature extraction module uses a convolutional neural network to extract image features of an affine image, image features of a translation-invariant image, and image features of a rotation-invariant image; An inverse transformation module, wherein the inverse transformation module performs an inverse Fourier transformation on the image features of the translation invariant image to obtain an inverse translation invariant image feature; and performs an inverse polar coordinate transformation on the image features of the rotation invariant image to obtain an inverse rotation invariant image feature; An inverse affine module, wherein the inverse affine module obtains a first image feature by performing an inverse affine rotation transformation on the image feature of the affine image, obtains a second image feature by performing an inverse affine rotation transformation on the inverse translation invariant image feature, and obtains a third image feature by performing an inverse affine rotation transformation on the inverse rotation invariant image feature; Feature fusion module, the feature fusion module Perform feature fusion on group image features to obtain fusion features; An acquisition module, wherein the acquisition module obtains an attention fusion feature by performing attention calculation on the fusion feature; By upsampling and normalizing the attention fusion features, rotation invariant features are obtained.