A rotation-invariant feature point detection and analysis method

Abstract

The present invention belongs to the technical field of computer digital image processing, and relates to a rotationally invariant feature point detection and analysis method. The present invention uses a deep learning model to build a feature point detection network, and adds an equivariant controllable backbone module, a global feature processing module, and a multi-resolution supervision module, and then trains a feature point detection system in a feature point detection network based on a drone map control point data set, and then when a remote sensing image of a map control feature point is input, the position of the map control feature point can be quickly obtained, saving manpower and material resources, and by judging the re-training requirements of the feature point detection system, it is helpful to improve the detection performance and robustness of the feature point detection system, and by analyzing the feature point position detection evaluation index of the feature point detection system for the target area, it is helpful to verify the accuracy of the feature point detection system, which is conducive to the automatic processing of a large number of image pictures.

Description

Technical Field

[0001] The invention belongs to the technical field of computer digital image processing and relates to a rotationally invariant feature point detection and analysis method. Background Art

[0002] The detection of ground control points is an important step in the use of drones. However, the detection of ground control points is currently mainly done manually, which makes it a task that consumes a lot of manpower and time. Therefore, a feature point detection method based on equivariant controllable convolution and geometrically constrained Gaussian heat map is of great significance and role.

[0003] In the prior art, there are also some related solutions for image feature point detection, for example, the invention patent application for image key point detection method, device, terminal and storage medium with Chinese patent publication number CN113838134B, which includes obtaining a target image and a preset feature map; performing convolution processing on the target image to obtain an output matrix, the output matrix is ​​composed of a plurality of matrix elements, wherein the matrix elements correspond to the feature map pixels in the preset feature map one by one; weighting the coordinates of the feature map pixels according to the matrix elements to obtain the weighted pixel coordinates of the feature map pixels corresponding to the matrix elements; summing the weighted pixel coordinates to obtain the coordinates of the key points in the preset feature map; determining the coordinates of the key points in the preset feature map mapped to the coordinates of the key points in the target image. In the embodiment of the present application, a new image key point detection method is provided, and the present application obtains the coordinates of the key points of the target image with only a small amount of calculation, which improves the efficiency of obtaining the coordinates of the key points in the target image and the coordinate accuracy.

[0004] Although the above schemes propose some solutions for image feature point detection, they still have certain limitations: on the one hand, the existing schemes mainly improve the efficiency and accuracy of obtaining the coordinates of the key points in the target image, but the existing schemes lack the use of learning models for learning and detecting image feature points, and thus cannot use neural networks to quickly and accurately detect the positions of image feature points.

[0005] On the other hand, the existing solutions ignore the verification of image feature point detection, so that the accuracy of image feature point detection cannot be verified, and further, the detection accuracy cannot be continuously optimized and improved, which is not conducive to improving processing efficiency. Summary of the invention

[0006] In view of this, in order to solve the problems raised in the above background technology, a rotation-invariant feature point detection and analysis method is now proposed.

[0007] The purpose of the present invention can be achieved through the following technical solutions: The present invention provides a rotation-invariant feature point detection and analysis method, including: S1, data set construction: obtaining each UAV remote sensing image, and marking control points therein, and then recording each UAV remote sensing image containing the marked control points as a target UAV map control point data set.

[0008] S2. Acquisition of feature point detection system: Use the deep learning model to build the feature point detection network to obtain the target feature point detection network, and add the equivariant controllable backbone module, global feature processing module and multi-resolution supervision module to it, and then train the feature point detection system in the target feature point detection network based on the target UAV map control point data set.

[0009] The content of the training feature point detection system in S2 includes: S21, based on the equivariant controllable backbone module, extracting the feature map with rotation invariant characteristics of each UAV remote sensing image, and recording it as the rotation invariant feature map of each UAV remote sensing image.

[0010] S22. Based on the global feature processing module, global feature modeling is performed on the rotation invariant feature map of each UAV remote sensing image to obtain a modeling feature map of each UAV remote sensing image.

[0011] S23. Based on the multi-resolution supervision module, the Gaussian heat map of each UAV remote sensing image is obtained.

[0012] S3, feature point detection system indicator analysis: obtain the accuracy, precision and recall rate of the feature point detection system, and determine the need for retraining of the feature point detection system. If retraining is required, execute S1; if retraining is not required, execute S4.

[0013] S4. Acquisition of feature point positions: Based on the Gaussian heat map of each UAV remote sensing image, the positions of each map control point in each UAV remote sensing image are calculated and recorded as the positions of each feature point in each UAV remote sensing image.

[0014] S5. Detection by feature point detection system: extract the target area corresponding to each sub-area image with known feature point positions from the database, record it as the target area corresponding to each sub-area image, perform detection and analysis on it by using the feature point detection system, obtain the position of each detected feature point in the target area corresponding to each sub-area image, and analyze the feature point position detection evaluation index of the feature point detection system for the target area.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention uses a deep learning model to build a feature point detection network, and adds an equivariant controllable backbone module, a global feature processing module and a multi-resolution supervision module, and then trains a feature point detection system in the feature point detection network based on the target UAV map control point data set. When the remote sensing image of the map control feature point is input, the position of the map control feature point can be quickly obtained, saving manpower and material resources.

[0016] 2. The present invention obtains the accuracy, precision and recall rate of the feature point detection system and determines the need for retraining of the feature point detection system, which helps to improve the detection performance and robustness of the feature point detection system.

[0017] 3. The present invention obtains the position of each detected feature point in the image of each sub-area corresponding to the target area, analyzes the feature point position detection evaluation index of the feature point detection system for the target area, helps to verify the accuracy of the feature point detection system, and improves the detection accuracy through continuous optimization, which is conducive to realizing the automated processing of a large number of image pictures, improving processing efficiency, reducing manual intervention, and reducing labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 It is a schematic diagram for implementing the scheme steps of the present invention.

[0020] Figure 2 It is a schematic diagram of the overall framework of the feature point detection system of the present invention.

[0021] Figure 3 It is a schematic diagram of the structure of the multi-resolution supervision module of the present invention. DETAILED DESCRIPTION

[0022] 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.

[0023] See also Figure 1As shown, the present invention provides a rotation-invariant feature point detection and analysis method, and the specific steps are as follows: S1, data set construction: obtain each UAV remote sensing image, and mark the control points therein, and then record each UAV remote sensing image containing the marked control points as the target UAV map control point data set.

[0024] S2. Acquisition of feature point detection system: Use the deep learning model to build the feature point detection network to obtain the target feature point detection network, and add the equivariant controllable backbone module, global feature processing module and multi-resolution supervision module to it, and then train the feature point detection system in the target feature point detection network based on the target UAV map control point data set.

[0025] It should be noted that the overall framework diagram of the feature point detection system is as follows Figure 2 shown.

[0026] As a preferred feasible embodiment, the specific operation of building a feature point detection network using a deep learning model includes: selecting a basic network, and adding a layer specifically used for feature point detection after the basic network, thereby completing the construction of the feature point detection network.

[0027] It should be further explained that the basic network includes but is not limited to VGGNet and ResNet.

[0028] The layers specifically used for feature point detection include but are not limited to fully connected layers and convolutional layers.

[0029] The content of the training feature point detection system in S2 includes: S21, based on the equivariant controllable backbone module, extracting the feature map with rotation invariant characteristics of each UAV remote sensing image, and recording it as the rotation invariant feature map of each UAV remote sensing image.

[0030] As a preferred feasible embodiment, the rotation invariant feature map of each UAV remote sensing image, and its specific acquisition process includes: A1, input each UAV remote sensing image, obtain the size of each UAV remote sensing image, use 3x3 equivariant controllable convolution on it, and obtain each UAV remote sensing image with a resolution reduced to half of the input, which is recorded as each first convolution UAV remote sensing image.

[0031] In a specific embodiment, assuming that the input is of size A UAV remote sensing image is obtained by using a 3x3 equivariant controllable convolution kernel to obtain a UAV remote sensing image with a resolution reduced to half of the input, that is, , where the stride of the equivariant controllable convolution kernel is 2, the padding size is 1, and the ReLU activation function is used.

[0032] It should be noted that the step size determines the number of pixels that the convolution kernel moves each time. A step size of 2 means that the convolution kernel moves once every two times, so the spatial dimension of the output will be reduced. For example, if the input is a 10x10 feature map, after using a 3x3 convolution kernel with a step size of 2, without considering padding, the output will be a (10-3) / 2+1=4x4 feature map.

[0033] The padding is to add an extra layer of zero values ​​at the edge of the input data to control the output size or ensure that the edge information is not lost. A padding size of 1 means adding a row / column of zero values ​​at each edge (up, down, left, and right) of the input data. The purpose of this is to allow the convolution kernel to cover the boundary part of the input data, thereby keeping the output size the same as the input or reducing it less.

[0034] The ReLU activation function sets all values ​​less than 0 to 0, and keeps values ​​greater than 0 unchanged. The introduction of the activation function can increase the nonlinear characteristics of the network. Through this simple truncation method, ReLU enables the network to learn a more complex and more realistic feature mapping relationship, and enhances the network's ability to distinguish and represent different feature patterns.

[0035] A2. Input each first convolution UAV remote sensing image, use a 3x3 equivariant controllable convolution kernel, and obtain each first convolution UAV remote sensing image with a resolution reduced to half of the input, which is recorded as each second convolution UAV remote sensing image.

[0036] In a specific embodiment, the once convolved UAV remote sensing image is again subjected to a 3x3 equivariant controllable convolution kernel to reduce the resolution of the once convolved UAV remote sensing image to 1 / 4 of the original image. ,Right now , and obtain the secondary convolution UAV remote sensing image.

[0037] A3. Input each secondary convolution UAV remote sensing image into the high-resolution preservation network to obtain the rotation invariant feature map of each UAV remote sensing image.

[0038] It needs to be further explained that the high-resolution preserving network is composed of downsampling and equivariant controllable convolution. Its core purpose is to be able to effectively integrate and abstract features like conventional operations in the process of in-depth processing of features, while maintaining a relatively high resolution of the feature map. Usually in many network architectures, with the increase in the number of network layers and continuous downsampling, the resolution of the feature map will become lower and lower. Although higher-level abstract features can be obtained, a lot of detail information will also be lost. The high-resolution preserving network here balances the needs of feature abstraction and detail retention to a certain extent through a specific structural design.

[0039] S22. Based on the global feature processing module, global feature modeling is performed on the rotation invariant feature map of each UAV remote sensing image to obtain a modeling feature map of each UAV remote sensing image.

[0040] As a preferred feasible embodiment, the specific acquisition process of the modeling feature map of each UAV remote sensing image includes: B1, expanding the rotation invariant feature map of each UAV remote sensing image in the last dimension.

[0041] It needs to be further explained that expanding the rotation invariant feature map of each UAV remote sensing image in the last dimension is to convert the 3D rotation invariant feature map of each UAV remote sensing image into a 2D form, that is, flattening the spatial dimensions (H, W) while retaining the channel information.

[0042] In a specific embodiment, the size of the rotation invariant feature map of a UAV remote sensing image is (32, 32, 64), wherein the height and width of the rotation invariant feature map of the UAV remote sensing image are 32 respectively, and the number of channels is 64. The rotation invariant feature map of the UAV remote sensing image is converted into a 2D form, and the size of the rotation invariant feature map of the UAV remote sensing image in 2D form is (32*32, 64), that is, the rotation invariant feature map of the UAV remote sensing image is flattened into a vector according to the channel dimension.

[0043] B2. Sample the last dimension of the rotation invariant feature map of each UAV remote sensing image processed by B1 to obtain a set size.

[0044] It should be further explained that the dimension sampling is usually for the purpose of reducing feature dimensions, reducing the amount of calculation or extracting key features. By sampling the last dimension of the rotation invariant feature map of each UAV remote sensing image processed by B1 to a set specific size, the features can be effectively simplified, making the feature representation more compact, while also avoiding the impact of excessive redundant information on subsequent processing, and helping to control the complexity of the model and the consumption of computing resources.

[0045] A specific embodiment, continuing the above embodiment, if the set size is set to 32, a dimensionality reduction method such as random projection can be used to implement it, thereby defining a projection matrix (64,32), and multiplying the size (32*32,64) of the rotation-invariant feature map of the UAV remote sensing image in 2D form by the projection matrix to obtain a new 2D rotation-invariant feature map of the UAV remote sensing image, whose size is (32*32,32), completing the sampling operation of the last dimension, and reducing the dimension of the feature as required.

[0046] B3. Add the rotation invariant feature map of each UAV remote sensing image processed by B2 to position embedding.

[0047] It should be further explained that the role of the position embedding is to incorporate position information into each element in the rotation invariant feature map of each UAV remote sensing image.

[0048] B4. Use multi-head self-attention to perform global attention modeling on the rotation-invariant feature map of each UAV remote sensing image processed by B3 to obtain the modeled feature map of each UAV remote sensing image.

[0049] It needs to be further explained that the multi-head self-attention mechanism enables the model to capture the relationship between features from different representation subspaces, that is, it can pay attention to the correlation between different parts of the feature map at the same time, perform global attention modeling, and help to explore the deep semantic information within the features and enhance the model's understanding and processing capabilities of complex patterns.

[0050] S23. Based on the multi-resolution supervision module, the Gaussian heat map of each UAV remote sensing image is obtained.

[0051] It should be noted that the schematic diagram of the multi-resolution supervision module is as follows Figure 3 shown.

[0052] It needs to be further explained that the addition of the multi-resolution supervision module is intended to enable the feature point detection network to eventually output a heat map with the same resolution as the input image. During the operation of the network, different layers will output feature maps of different resolutions. This module will use these multi-resolution feature maps to generate corresponding predicted heat maps for them respectively, and define a loss function for the predicted heat map at each resolution, compare it with the real annotated heat map, and use back propagation to optimize the network parameters so that the predicted heat map at each resolution continues to approach the actual situation. After that, the predicted heat maps of different resolutions are integrated through a suitable fusion method, and finally a heat map with the same resolution as the input image is generated. Such a heat map can more finely and accurately reflect the probability of feature points appearing at various positions in the original image, which plays a key role in accurately detecting the position of feature points.

[0053] In a specific example, the loss function is a mean square error loss function.

[0054] The fusion method is, for example, upsampling combined with weighted fusion.

[0055] As a preferred feasible embodiment, the Gaussian heat map of each UAV remote sensing image is specifically acquired through the following steps: C1, using 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image to 6.

[0056] It needs to be further explained that the 1x1 convolution plays a unique role in deep learning networks. Although its convolution kernel size is very small and only involves operations around one pixel, it can effectively change the number of channels of the feature map. From a mathematical point of view, for an input feature map (assuming the shape is [batch_size, height, width, in_channels], batch_size represents the number of samples processed at a time, height and width are the height and width of the feature map, and in_channels is the number of input channels), after 1x1 convolution, it is actually the input channel data corresponding to each pixel and the 1x1 convolution kernel (whose number determines the number of output channels, here set to output 6 channels) The weighted sum operation is performed, thereby converting the input feature map into a new feature map with a specified number of output channels (the shape becomes [batch_size, height, width, 6]).

[0057] C2. Use deconvolution to change the resolution of the modeling feature map of each UAV remote sensing image, so that its resolution is doubled, and the modeling feature map of each UAV remote sensing image with enlarged resolution is obtained. Then use 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image with enlarged resolution to 6.

[0058] It should be further explained that the deconvolution is also called transposed convolution or inverse convolution, which is an operation used to improve the resolution of feature maps. Its function is opposite to that of conventional convolution. Conventional convolution reduces the size of feature maps and aggregates features by downsampling, while deconvolution can enlarge the size of feature maps according to set parameters and restore some detail information.

[0059] The specific implementation method of doubling the resolution of the modeling feature map of each UAV remote sensing image is: by adjusting the deconvolution parameters, the number of pixels of the feature map in the horizontal and vertical directions is doubled, thereby obtaining a feature map of a larger size.

[0060] Among them, the parameters of deconvolution include but are not limited to step size, convolution kernel size and padding.

[0061] The purpose of using 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image after the resolution is enlarged to 6 is to further adjust the representation of the features to meet the subsequent processing or task requirements.

[0062] C3. Use deconvolution to change the resolution of the modeling feature map of each UAV remote sensing image after processing by C2, so that its resolution is doubled, and the modeling feature map of each UAV remote sensing image with enlarged resolution after processing by C2 is obtained. Then use 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image with enlarged resolution after processing by C2 to 6, and then record the modeling feature map of each UAV remote sensing image with enlarged resolution after processing by C2 that has been finally processed as the Gaussian heat map of each UAV remote sensing image.

[0063] The present invention uses a deep learning model to build a feature point detection network, and adds an equivariant controllable backbone module, a global feature processing module and a multi-resolution supervision module, and then trains a feature point detection system in the feature point detection network based on a target UAV map control point data set. When a remote sensing image of a map control feature point is input, the position of the map control feature point can be quickly obtained, saving manpower and material resources.

[0064] S3, feature point detection system indicator analysis: obtain the accuracy, precision and recall rate of the feature point detection system, and determine the need for retraining of the feature point detection system. If retraining is required, execute S1; if retraining is not required, execute S4.

[0065] It should be further explained that the specific method of obtaining the accuracy, precision and recall rate of the feature point detection system is: directly obtain the true positive examples, true negative examples, false positive examples and false negative examples of the feature point detection system from the output results of the feature point detection system, and record them as , and then based on the standard calculation formula, the accuracy of the feature point detection system can be obtained respectively , Accuracy and recall .

[0066] It should be noted that the true positive examples are the number of positive samples that are correctly judged as positive classes.

[0067] The true negative examples are the number of negative samples that are correctly judged as negative classes.

[0068] The false positive examples are the number of negative samples that are mistakenly judged as positive.

[0069] The false negative examples are the number of positive samples that are mistakenly judged as negative samples.

[0070] Among them, positive samples are feature points that exist, and negative samples are feature points that are considered to exist.

[0071] It should be further explained that the specific calculation formulas for the accuracy, precision and recall rate of the feature point detection system can be obtained according to the standard calculation formula: Get the accuracy of the feature point detection system Accuracy is a more intuitive indicator, but it may be affected by data imbalance. It reflects the recognition effect of feature points.

[0072] According to the calculation formula Get the accuracy of the feature point detection system The precision rate can measure the accuracy of the feature point detection system in identifying feature points.

[0073] According to the calculation formula Get the recall rate of the feature point detection system The recall rate can measure the ability of the feature point detection system to correctly identify feature points.

[0074] As a preferred feasible embodiment, in the process of determining the retraining demand of the feature point detection system, it is necessary to construct a retraining demand evaluation coefficient of the feature point detection system. The specific analysis method includes: extracting the accuracy, precision and recall rate of the feature point detection system, which are respectively recorded as .

[0075] Analysis of feature point detection system value .

[0076] According to the analytical formula Get the retraining requirement evaluation coefficient of the feature point detection system ,in are the accuracy threshold and Value threshold, is a natural constant.

[0077] The present invention obtains the accuracy, precision and recall rate of the feature point detection system and determines the need for retraining of the feature point detection system, which helps to improve the detection performance and robustness of the feature point detection system.

[0078] S4. Acquisition of feature point positions: Based on the Gaussian heat map of each UAV remote sensing image, the positions of each map control point in each UAV remote sensing image are calculated and recorded as the positions of each feature point in each UAV remote sensing image.

[0079] As a preferred feasible embodiment, the position of each map control point in each UAV remote sensing image is specifically obtained in the following manner: extracting the Gaussian heat map of each UAV remote sensing image, sequentially accessing the Gaussian heat map corresponding to each pixel point of each UAV remote sensing image through a direct traversal search method, and comparing the probability values, so as to obtain the position of each pixel point corresponding to the Gaussian heat map of each UAV remote sensing image with the largest probability value, and record it as the peak position corresponding to the Gaussian heat map of each UAV remote sensing image, that is, the position of each map control point in each UAV remote sensing image.

[0080] It needs to be further explained that the direct traversal search method uses a programming language to sequentially access and compare the probability values ​​of each pixel point corresponding to the Gaussian heat map of each drone remote sensing image.

[0081] Among them, programming languages ​​include but are not limited to Python, Java and C++.

[0082] In a specific embodiment, the specific process of obtaining the peak positions of the Gaussian heat map of each UAV remote sensing image using Python language is as follows: (1) Import the numpy library and create a simple example heatmap.

[0083] (2) Initialize max_value and use it as the probability value of the first pixel in the upper left corner of the heat map. At the same time, initialize the row index max_row and column index max_col corresponding to the maximum value to 0.

[0084] (3) Use two nested for loops, with the outer loop controlling the row index row and the inner loop controlling the column index col, to traverse the entire heat map array. During the loop, each time the probability value of the current pixel point heatmap[row,col] is compared with the recorded maximum value max_value. If the current value is larger, then max_value, max_row, and max_col are updated.

[0085] (4) Output the location (row and column index) where the probability value is the largest and the maximum value itself.

[0086] S5. Detection by feature point detection system: extract the target area corresponding to each sub-area image with known feature point positions from the database, record it as the target area corresponding to each sub-area image, perform detection and analysis on it by using the feature point detection system, obtain the position of each detected feature point in the target area corresponding to each sub-area image, and analyze the feature point position detection evaluation index of the feature point detection system for the target area.

[0087] As a preferred feasible embodiment, the feature point detection system detects and evaluates the feature point position of the target area, and its specific analysis method includes: extracting the position of each detected feature point in the image picture of each sub-area corresponding to the target area, and extracting the position of each standard feature point in the image picture of each sub-area corresponding to the target area from the database.

[0088] The lower boundary of the target area corresponding to each sub-area image is used as the horizontal coordinate, the left boundary is used as the vertical coordinate, and the intersection of the lower boundary and the left boundary is used as the coordinate origin to establish a two-dimensional coordinate system, and then the coordinate points of the positions of the detected feature points in the target area corresponding to each sub-area image are obtained. And the coordinate points of each standard feature point position in the target area corresponding to each sub-area image picture ,in , is the number of each sub-region, is the number of sub-regions, , is the number of each detected feature point, is the number of detected feature points, , is the number of each standard feature point, is the number of standard feature points.

[0089] Determine the number of overlaps between the coordinate points of each detected feature point in the target area corresponding to each sub-area image and the coordinate points of each standard feature point in the target area corresponding to each sub-area image, and record it as the number of successful feature point detections in the target area corresponding to each sub-area image , analyze the feature point detection system for the target area feature point position detection evaluation index ,in is the total number of standard feature points of the target area extracted from the database.

[0090] As a preferred feasible embodiment, the method uses a database during execution to store the accuracy threshold and The value threshold is set, and the image pictures of each sub-region corresponding to the target area with known feature point positions are stored, the positions of each standard feature point in the image pictures of each sub-region corresponding to the target area are stored, and the total number of standard feature points in the target area is stored.

[0091] The present invention obtains the position of each detected feature point in the image pictures of each sub-area corresponding to the target area, analyzes the feature point position detection evaluation index of the feature point detection system for the target area, helps to verify the accuracy of the feature point detection system, and improves the detection accuracy through continuous optimization, which is conducive to realizing the automated processing of a large number of image pictures, improving processing efficiency, reducing manual intervention, and reducing labor costs.

[0092] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they shall all fall within the protection scope of the present invention.

Claims

1. A rotation-invariant feature point detection and analysis method, characterized in that: include: S1. Dataset construction: Obtain each UAV remote sensing image and mark the control points in it, and then record each UAV remote sensing image containing the marked control points as the target UAV map control point dataset; S2. Acquisition of feature point detection system: Use the deep learning model to build a feature point detection network to obtain a target feature point detection network, and add an equivariant controllable backbone module, a global feature processing module, and a multi-resolution supervision module to it. Then, based on the target UAV map control point data set, the target feature point detection network is trained to obtain a feature point detection system; The training feature point detection system in S2 includes: S21, based on the equivariant controllable backbone module, extracting the feature map with rotation invariant characteristics of each UAV remote sensing image, and recording it as the rotation invariant feature map of each UAV remote sensing image; S22, based on the global feature processing module, performing global feature modeling on the rotation invariant feature map of each UAV remote sensing image to obtain a modeling feature map of each UAV remote sensing image; S23, based on the multi-resolution supervision module, processing the modeling feature map of each UAV remote sensing image to obtain a Gaussian heat map of each UAV remote sensing image; S3, feature point detection system indicator analysis: obtain the accuracy, precision and recall rate of the feature point detection system, and determine the need for retraining of the feature point detection system. If retraining is required, execute S1; if retraining is not required, execute S4; S4, feature point position acquisition: based on the Gaussian heat map of each UAV remote sensing image, the position of each map control point in each UAV remote sensing image is calculated and recorded as the position of each feature point in each UAV remote sensing image; S5. Detection by feature point detection system: extract the target area corresponding to each sub-area image with known feature point positions from the database, record it as the target area corresponding to each sub-area image, perform detection and analysis on it by using the feature point detection system, obtain the position of each detected feature point in the target area corresponding to each sub-area image, and analyze the feature point position detection evaluation index of the feature point detection system for the target area.

2. The rotation-invariant feature point detection and analysis method according to claim 1, characterized in that: The specific operations of building a feature point detection network using a deep learning model include: selecting a basic network, and adding a layer specifically used for feature point detection after the basic network, thereby completing the construction of the feature point detection network.

3. The rotation-invariant feature point detection and analysis method according to claim 1, characterized in that: The specific acquisition process of the rotation invariant feature map of each UAV remote sensing image includes: A1. Input each UAV remote sensing image, obtain the size of each UAV remote sensing image, apply 3x3 equivariant controllable convolution to it, and obtain each UAV remote sensing image with a resolution reduced to half of the input, which is recorded as each first convolution UAV remote sensing image; A2. Input each first convolution UAV remote sensing image, use a 3x3 equivariant controllable convolution kernel, and obtain each first convolution UAV remote sensing image with a resolution reduced to half of the input, which is recorded as each second convolution UAV remote sensing image; A3. Input each secondary convolution UAV remote sensing image into the high-resolution preservation network to obtain the rotation invariant feature map of each UAV remote sensing image.

4. The method for detecting and analyzing rotationally invariant feature points according to claim 3, characterized in that: The specific acquisition process of the modeling feature graph of each UAV remote sensing image includes: B1. Expand the rotation invariant feature map of each UAV remote sensing image in the last dimension; B2, sampling the last dimension of the rotation invariant feature map of each UAV remote sensing image processed by B1 to obtain a set size; B3, adding the rotation invariant feature map of each UAV remote sensing image processed by B2 to position embedding; B4. Use multi-head self-attention to perform global attention modeling on the rotation-invariant feature map of each UAV remote sensing image processed by B3 to obtain the modeled feature map of each UAV remote sensing image.

5. The method for detecting and analyzing rotationally invariant feature points according to claim 4, characterized in that: The specific acquisition process of the Gaussian heat map of each UAV remote sensing image includes: C1, use 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image to 6; C2. Use the deconvolution convolution to change the resolution of the modeling feature map of each UAV remote sensing image, so that its resolution is doubled, and the modeling feature map of each UAV remote sensing image with enlarged resolution is obtained. Then use 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image with enlarged resolution to 6; C3. Use deconvolution to change the resolution of the modeling feature map of each UAV remote sensing image after processing by C2, so that its resolution is doubled, and the modeling feature map of each UAV remote sensing image with enlarged resolution after processing by C2 is obtained. Then use 1x1 convolution to change the number of channels of the modeling feature map of each UAV remote sensing image with enlarged resolution after processing by C2 to 6, and then record the modeling feature map of each UAV remote sensing image with enlarged resolution after processing by C2 that has been finally processed as the Gaussian heat map of each UAV remote sensing image.

6. The method for detecting and analyzing rotationally invariant feature points according to claim 1, characterized in that: In the process of determining the retraining demand of the feature point detection system, it is necessary to construct a retraining demand evaluation coefficient of the feature point detection system, and the specific analysis method includes: The accuracy, precision and recall of the feature point detection system are extracted and recorded as ; Analysis of feature point detection system value ; According to the analytical formula Get the retraining requirement evaluation coefficient of the feature point detection system ,in are the accuracy threshold and Value threshold, is a natural constant.

7. The method for detecting and analyzing rotationally invariant feature points according to claim 6, characterized in that: The retraining requirement of the feature point detection system is specifically determined in the following manner: The re-training need evaluation coefficient of the feature point detection system is compared with the set re-training need evaluation coefficient threshold. If the re-training need evaluation coefficient of the feature point detection system is greater than or equal to the re-training need evaluation coefficient threshold, the re-training need of the feature point detection system is recorded as requiring re-training; otherwise, the re-training need of the feature point detection system is recorded as not requiring re-training.

8. The method for detecting and analyzing rotationally invariant feature points according to claim 1, characterized in that: The specific method of obtaining the position of each image control point in each UAV remote sensing image includes: The Gaussian heat map of each UAV remote sensing image is extracted, and the corresponding pixel points of the Gaussian heat map of each UAV remote sensing image are visited in turn through the direct traversal search method, and their probability values ​​are compared to obtain the Gaussian heat map of each UAV remote sensing image with the largest probability value. The corresponding pixel point position is recorded as the peak position of the Gaussian heat map of each UAV remote sensing image, which is the position of each map control point in each UAV remote sensing image.

9. The method for detecting and analyzing rotationally invariant feature points according to claim 1, characterized in that: The feature point detection system detects the evaluation index of the feature point position of the target area, and the specific analysis method includes: Extract the position of each detected feature point in the image picture of each sub-region corresponding to the target area, and extract the position of each standard feature point in the image picture of each sub-region corresponding to the target area from the database; The lower boundary of the target area corresponding to each sub-area image is used as the horizontal coordinate, the left boundary is used as the vertical coordinate, and the intersection of the lower boundary and the left boundary is used as the coordinate origin to establish a two-dimensional coordinate system, and then the coordinate points of the positions of the detected feature points in the target area corresponding to each sub-area image are obtained. And the coordinate points of each standard feature point position in the target area corresponding to each sub-area image picture ,in , is the number of each sub-region, is the number of sub-regions, , is the number of each detected feature point, is the number of detected feature points, , is the number of each standard feature point, is the number of standard feature points; Determine the number of overlaps between the coordinate points of each detected feature point in the target area corresponding to each sub-area image and the coordinate points of each standard feature point in the target area corresponding to each sub-area image, and record it as the number of successful feature point detections in the target area corresponding to each sub-area image , analyze the feature point detection system for the target area feature point position detection evaluation index ,in is the total number of standard feature points of the target area extracted from the database.

10. The method for detecting and analyzing rotationally invariant feature points according to claim 1, characterized in that: The method uses a database during execution to store the accuracy threshold and The value threshold is set, and the image pictures of each sub-region corresponding to the target area with known feature point positions are stored, the positions of each standard feature point in the image pictures of each sub-region corresponding to the target area are stored, and the total number of standard feature points in the target area is stored.

Images