Method for ground object extraction based on remote sensing images

By combining the fusion of spectral, texture, context and shape features in remote sensing image processing, a vector topographic network is generated, which solves the problem of insufficient accuracy and automation of ground object extraction in the prior art, and achieves more efficient calculation and more accurate identification .

CN118570596BActive Publication Date: 2024-11-29ZHONGKE SHANSHUI (BEIJING) TECH INFORMATION CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410660312.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-27
Publication Date
2024-11-29
Estimated Expiration
2044-05-27

AI Technical Summary

Technical Problem

Existing remote sensing image processing technology is difficult to effectively combine spectral, texture, context and shape characteristics, resulting in insufficient accuracy and automation of ground objects and low computing efficiency.

Method used

By obtaining the spectrum, texture, context and shape characteristics of the remote sensing image, feature fusion is performed, a vector terrain network is generated, and ground objects are then extracted. The specific steps include: calculating the pixel density distribution, determining local maxima points, and serializing; fusing based on these features, generating a fusion feature set of spectral, texture, context and shape, and using these features to generate a vector terrain network.

Benefits of technology

It improves the accuracy and automation of ground object extraction, enhances the ability to identify different land objects types, reduces noise interference, improves calculation efficiency, and can more accurately identify land objects with similar spectral characteristics but different spatial distributions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118570596B_ABST
    Figure CN118570596B_ABST
Patent Text Reader

Abstract

In the method for extracting ground objects based on remote sensing images in the present application, feature extraction is performed on the remote sensing image to obtain spectral features, texture features, context features and shape features. By fusing the spectral features with the texture features, and fusing the context features with the shape features, two fused feature sets are generated to solve the redundancy and complementarity problems between features, enhance the feature expression capability, and improve the recognition accuracy of ground objects. The fused features are further used to generate a vector terrain network, convert the pixel information of the remote sensing image into vector information, more accurately represent the terrain features in the geographic space, and solve the problem of using context information and shape information. Finally, the ground objects on the remote sensing image are extracted based on the vector terrain network. The vector terrain network serves as a guide in the extraction process, which helps to identify and distinguish objects with similar spectral characteristics but different spatial distributions, thereby improving the automation and accuracy of the extraction process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent processing technology, and in particular to a method for extracting ground objects based on remote sensing images. Background Art

[0002] As an efficient means of obtaining geographic information, remote sensing technology plays an important role in agriculture, forestry, urban planning, environmental protection, disaster monitoring, etc. With the continuous enhancement of remote sensing data acquisition capabilities, the resolution of remote sensing images is getting higher and higher, and the amount of information is getting richer and richer, which makes it possible to automatically extract ground objects.

[0003] Traditional remote sensing image interpretation methods mainly rely on single image features, such as spectral features or texture features, to identify and classify ground objects. However, due to the complexity and diversity of ground objects, a single feature is often difficult to meet the needs of high-precision extraction. For example, different ground objects may have similar responses in the spectrum, while the same type of ground objects may have different texture features under different environmental conditions.

[0004] In addition, the contextual information and shape information of remote sensing images are also important for the recognition of ground objects. Contextual features can provide information about the relationship and spatial distribution between objects, while shape features help to identify objects with specific shapes. However, how to effectively combine these different types of features, and how to process and fuse the relationship between them, is a technical challenge facing the current field of remote sensing image processing.

[0005] In order to solve the above problems, researchers began to explore multi-feature fusion technology, hoping to improve the accuracy and robustness of ground object extraction by comprehensively utilizing multiple information in remote sensing images. Multi-feature fusion technology constructs a complex feature space, enabling the algorithm to understand and identify ground objects more comprehensively, thus overcoming the limitations of a single feature to a certain extent.

[0006] Although multi-feature fusion technology has made some progress, there are still some difficulties and challenges in practical applications. For example, how to determine the weights and fusion strategies of different features, how to deal with the redundancy and complementary relationship between features, and how to improve the computational efficiency and practicality of the algorithm are all issues that need further research and resolution.

[0007] In addition, with the continuous expansion of remote sensing application fields, the real-time and automation requirements for remote sensing image processing technology are becoming increasingly higher. Therefore, researching and developing more efficient, accurate, and automated ground object extraction methods has important theoretical and practical significance. Summary of the invention

[0008] In order to solve the above technical problems, the present application provides a method for extracting ground objects based on remote sensing images, so as to at least solve or alleviate the above problems existing in the prior art.

[0009] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a method for extracting ground objects based on remote sensing images is provided, which includes: acquiring a ground remote sensing image; performing feature extraction on the ground remote sensing image to obtain spectral features, texture features, context features, and shape features; fusing the spectral features and texture features to obtain a first fused feature; fusing the context features and shape features to obtain a second fused feature; generating a vector terrain network based on the first fused feature and the second fused feature; and extracting the ground objects on the remote sensing image based on the vector terrain network.

[0010] In the present application, feature extraction is performed on remote sensing images to obtain multiple features including spectral features, texture features, context features and shape features, which solves the problem mentioned in the prior art that a single feature is difficult to fully describe ground objects. By integrating multiple features, the algorithm's ability to recognize different types of objects is increased. In addition, by fusing spectral features with texture features, and fusing context features with shape features, two fused feature sets are generated. This fusion strategy aims to solve the redundancy and complementarity problems between features, enhance the feature expression ability, and improve the recognition accuracy of ground objects. Finally, using the above-mentioned fusion features, the scheme further generates a vector terrain network, converts the pixel information of the remote sensing image into vector information, helps to more accurately represent the terrain features in the geographic space, and solves the problem of using context information and shape information mentioned in the prior art. Finally, in the present application, ground objects on remote sensing images are extracted according to the vector terrain network. The vector terrain network, as a guide in the extraction process, helps to identify and distinguish objects with similar spectral characteristics but different spatial distributions, such as rivers, roads, etc., thereby improving the automation and accuracy of the extraction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The schematic embodiment drawings and their descriptions of this application are used to explain this application and do not constitute an improper limitation on this application. In the drawings:

[0012] Figure 1 The present invention is a flowchart of a method for extracting ground objects based on remote sensing images according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] In order to enable those skilled in the art to better understand the solution of the present application, the following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0014] Figure 1 FIG. 1 is a flow chart of a method for extracting ground objects based on remote sensing images according to an embodiment of the present application. Figure 1 As shown, it includes:

[0015] Acquisition of ground remote sensing images;

[0016] Extracting features from the ground remote sensing image to obtain spectral features, texture features, context features, and shape features;

[0017] Fusing the spectral feature and the texture feature to obtain a first fused feature;

[0018] Fusing the context feature and the shape feature to obtain a second fused feature;

[0019] generating a vector terrain network according to the first fused features and the second fused features;

[0020] The ground objects on the remote sensing image are extracted according to the vector terrain network.

[0021] In the present application, feature extraction is performed on remote sensing images to obtain multiple features including spectral features, texture features, context features and shape features, which solves the problem mentioned in the prior art that a single feature is difficult to fully describe ground objects. By integrating multiple features, the algorithm's ability to recognize different types of objects is increased. In addition, by fusing spectral features with texture features, and fusing context features with shape features, two fused feature sets are generated. This fusion strategy aims to solve the redundancy and complementarity problems between features, enhance the feature expression ability, and improve the recognition accuracy of ground objects. Finally, using the above-mentioned fusion features, the scheme further generates a vector terrain network, converts the pixel information of the remote sensing image into vector information, helps to more accurately represent the terrain features in the geographic space, and solves the problem of using context information and shape information mentioned in the prior art. Finally, in the present application, ground objects on remote sensing images are extracted according to the vector terrain network. The vector terrain network, as a guide in the extraction process, helps to identify and distinguish objects with similar spectral characteristics but different spatial distributions, such as rivers, roads, etc., thereby improving the automation and accuracy of the extraction process.

[0022] Optionally, the feature extraction of the ground remote sensing image to obtain spectral features includes: calculating the pixel density distribution on the ground remote sensing image; determining a plurality of pixel local maximum points on the ground remote sensing image based on the pixel density distribution; serializing the plurality of pixel local maximum points based on the positions of the pixel points to obtain a local maximum point sequence; and extracting features of the ground remote sensing image based on the local maximum point sequence to obtain spectral features.

[0023] To this end, the above solution has the following beneficial technical effects:

[0024] 1. More accurate spectral feature extraction: Calculating pixel density distribution can more accurately reflect the brightness or color changes of different types of objects on the image, which is the key to extracting spectral features. Determining the local maximum points of pixels helps to identify the "peaks" or "key points" representing different types of objects in the image, which usually correspond to the spectral responses of different objects.

[0025] 2. Improve feature extraction efficiency: Serializing multiple local maximum points of pixels can form an ordered set of feature points, which is convenient for subsequent processing and analysis. The ordered local maximum point sequence helps reduce redundant calculations and data processing time, and improves the efficiency of the entire feature extraction process.

[0026] 3. Enhance the robustness of the algorithm: Spectral feature extraction based on local maximum points can reduce the impact of noise and other interference factors on the feature extraction results, because local maximum points usually have good robustness to noise.

[0027] 4. Support the subsequent fusion process: The extracted spectral features will be fused with other features (such as texture features, context features, shape features) to provide a more comprehensive and accurate description of ground objects. Accurate spectral feature extraction can provide more valuable input for the subsequent fusion process, thereby enhancing the recognition ability of the entire system.

[0028] 5. Improve the accuracy of ground object extraction: Accurate spectral features can help the algorithm better distinguish between objects with similar spectral characteristics but different spatial distributions, such as rivers, roads, etc. By accurately extracting and fusing spectral features, the automation and accuracy of ground object extraction can be improved, providing more valuable data support for subsequent geospatial analysis and applications.

[0029] This application provides an exemplary code as follows:

[0030]

[0031]

[0032] Optionally, calculating the pixel density distribution on the ground remote sensing image includes:

[0033] For any pixel point on the ground remote sensing image, obtain its pixel value, and determine the pixel values ​​of all pixel points in the neighborhood constructed for the any pixel point;

[0034] Based on the formula Calculate the pixel density, x, with any pixel point as reference i represents the pixel value of the i-th pixel on the ground remote sensing image, x j is the pixel value of any pixel in the neighborhood of the ith pixel, N is the number of pixels considered in the neighborhood, h is the bandwidth parameter used to determine the size of the neighborhood, K is the kernel function, f(x i ) is the pixel density of the ith pixel; based on the pixel density of all pixels as a reference, the pixel density gradient distribution of any pixel on the ground remote sensing image is calculated. To this end, the above scheme has the following beneficial technical effects: by calculating the pixel density distribution, the characteristics of different areas in the ground remote sensing image, such as terrain, vegetation, buildings, etc., can be more accurately described. Enhanced image analysis effect: The pixel density gradient distribution can provide information about the changes in pixel values ​​between different areas in the image, which is very useful for identifying features such as edges, textures and patterns in the image, thereby enhancing the effect of image analysis. The calculation of pixel density and gradient distribution is in remote sensing image processing, as well as adjusting the neighborhood size and pixel value calculation method. This method can adapt to images of different resolutions and qualities.

[0035] This application provides an exemplary code as follows:

[0036]

[0037] Optionally, the pixel density gradient distribution of any pixel point on the ground remote sensing image is calculated based on the pixel density of all pixel points referenced based on the following formula, including:

[0038]

[0039] Where N is the total number of pixels in the neighborhood, w j is the weight, K h (x i -x j ) is x i -x j Kernel function, h is the bandwidth parameter used for smoothing, f(x j ) and f(x i ) are the pixel densities of the j-th pixel and the i-th pixel, respectively. is the pixel density gradient distribution of the i-th pixel. To this end, by calculating the pixel density gradient distribution, the pixel value changes between different regions in the image can be more accurately depicted. This is crucial for analyzing local features of images, edge detection, and texture analysis. The gradient distribution provides detailed information about the direction and intensity of pixel value changes, which is crucial for understanding the spatial relationship between objects and backgrounds in images. By adjusting the kernel function and bandwidth parameters, it can adapt to different types of remote sensing images and data sets, enhancing the flexibility and adaptability of the technology and enabling it to be applied in a wide range of scenarios.

[0040] This application provides an exemplary code as follows:

[0041]

[0042]

[0043] The code above uses SciPy's filters.convolve function to apply a Gaussian kernel for smoothing, and uses NumPy's gradient function to calculate the gradient. The size of the kernel function (kernel_size) is determined by the bandwidth parameter (bandwidth) and the standard deviation (sigma). The gradient magnitude gradient_magnitude provides a comprehensive measure of the gradient distribution, while grad_x and grad_y provide the horizontal and vertical gradient components, respectively.

[0044] Optionally, determining a plurality of local maximum points of pixels on the ground remote sensing image according to the pixel density distribution includes:

[0045] Based on the set iteration step, according to the formula The pixel value x for the i-th pixel i Iterate, where t represents the current iteration step number, and t+1 represents the next iteration step number. Indicates the pixel value corresponding to the current iteration step, Represents the pixel value corresponding to the next iteration step, η is a positive step size parameter;

[0046] Calculate the pixel density update value corresponding to the corresponding pixel value at the next iteration step If satisfied Then the pixel density f(x i) is a local maximum point of a pixel, and ∈ is a set threshold. To this end, the above scheme has the following beneficial technical effects: local maximum points often correspond to significant features or targets in the image, such as mountains, buildings, road intersections, etc. By identifying these points, key features in the image can be automatically extracted and marked, providing important information for subsequent image analysis, target recognition or classification. In remote sensing images, a large number of pixels may contain redundant information. By identifying local maximum points, the number of pixels to be processed can be significantly reduced, thereby speeding up processing and reducing computational costs. The local maximum point detection method based on pixel density distribution has a certain robustness to noise and image changes. This is because the local maximum point is a point with a significantly larger value relative to its surrounding pixels in a certain neighborhood, which can resist a certain degree of noise interference and image changes. The local maximum point has a clear physical meaning and can intuitively explain the features and targets in the image.

[0047] This application provides an exemplary code as follows:

[0048]

[0049]

[0050] Optionally, the method of serializing the multiple local maximum points of pixels based on the positions of the pixels to obtain a local maximum point sequence includes: counting the feature intensity of each local maximum point of each pixel; and serializing the feature intensity of the multiple local maximum points of pixels based on the positions of the pixels to obtain a local maximum point sequence. To this end, the above scheme has the following beneficial technical effects: the serialization process is sorted according to the feature intensity, ensuring that the most important or most significant local maximum points in the image can be processed first, and that those points that are most likely to represent important information or targets can be given priority. In addition, since the serialization is based on the position of the pixels, the spatial relationship between the local maximum points is maintained, which helps to understand the relative position relationship between the structure, pattern or target in the image, and further improves the accuracy of image analysis.

[0051] This application provides an exemplary code as follows:

[0052]

[0053]

[0054] Optionally, the statistical feature strength of each pixel local maximum point includes: calculating the ratio of the pixel density of each pixel local maximum point to the mean of the pixel density in its neighborhood; calculating the variance of the pixel density of each pixel local maximum point and the pixel density in its neighborhood; based on the formula A weighted operation is performed on the ratio and the variance to calculate the characteristic intensity of the local maximum point of each pixel, where I i is the characteristic intensity of the local maximum point of each pixel, R i is the ratio of each pixel’s local maximum point to the mean pixel density in its neighborhood. is the variance of the pixel density in the neighborhood, and α and β are weight parameters used to balance the contribution of ratio and variance to feature intensity. To this end, the above scheme has the following beneficial technical effects: Local contrast enhancement: Calculating the ratio of the local maximum point of a pixel to the mean value of the pixel density in its neighborhood can highlight the contrast between the point and the surrounding area. This helps to identify those points that are significantly prominent in the local area, such as edges, corners, etc., which usually contain important structural or texture information in the image. Texture and detail capture: The variance of the pixel density in the neighborhood can reflect the degree of discreteness of the pixel values ​​in the area, that is, the richness of the texture and the amount of detail information. By calculating the variance and incorporating it into the calculation of the feature intensity, the ability to capture texture and detail information in the image can be enhanced. Adjustability: Introducing weight parameters and allowing users to adjust the contribution of ratio and variance in feature intensity calculation according to specific application requirements. By adjusting these parameters, the sensitivity of feature intensity to local contrast and texture details can be flexibly controlled, thereby achieving optimized processing of different image features. This method combines local contrast and texture information, so that the feature intensity has a certain robustness to interference factors such as noise and illumination changes. By comprehensively considering multiple factors, this method can more accurately describe the characteristics of local maximum points of pixels and improve the accuracy and reliability of image analysis and processing. Since it only involves the calculation of pixel density and pixel density in the neighborhood, this method is highly efficient when processing large-scale image data.

[0055] This application provides an exemplary code as follows:

[0056]

[0057]

[0058] Optionally, the method of serializing the characteristic intensities of the plurality of local maximum points of pixels based on the positions of the pixels to obtain a sequence of local maximum points includes: constructing a characteristic vector based on the positions of the pixels and the corresponding local maximum points of the pixels at the positions; collecting the characteristic vectors of all local maximum points into a matrix, wherein each row represents a vector of a local maximum point, to obtain a principal component analysis matrix, wherein the principal component analysis matrix has zero mean and unit variance; calculating the eigenvalues ​​of the principal component analysis matrix, to serialize the plurality of local maximum points of pixels based on the eigenvalues, to obtain a sequence of local maximum points. To this end, the above scheme has the following beneficial technical effects: through principal component analysis, high-dimensional characteristic vectors can be projected into a low-dimensional space while maintaining the main information. This helps to reduce the dimension of the data, reduce the complexity of calculation, and simplify subsequent data processing and analysis steps. Principal component analysis can eliminate redundant information in the characteristic vector, i.e., those features that contribute less to describing data changes. This helps to highlight the main components in the image and improve the efficiency and accuracy of image analysis. By calculating the eigenvalues ​​of the principal component analysis matrix, different local maximum points can be sorted and selected. Points with larger eigenvalues ​​usually correspond to more important features in the data, so we can selectively focus on or ignore certain local maximum points based on the size of the eigenvalues. The sequence of local maximum points obtained by principal component analysis has a certain order, which helps to perform operations according to a certain logic or sequence in subsequent processing. We can merge, segment or track regions based on the sequence of local maximum points. Since PCA analysis takes into account the position information of pixels, the sequence of local maximum points obtained has a certain robustness to image translation, rotation and other transformations, improving the stability and generalization ability of image processing algorithms.

[0059] This application provides an exemplary code as follows:

[0060]

[0061]

[0062] In the code above, the feature strength is a value directly associated with each local maximum point. The location information (x and y coordinates) is also converted into part of the feature vector. Then, sklearn.decomposition.PCA is used to perform PCA analysis and the results are stored in pca_features. Each row in this array corresponds to a PCA feature vector of a local maximum point, so it can be regarded as a serialized sequence of local maximum points. Finally, the feature vectors are sorted according to the first principal component score of PCA.

[0063] Optionally, the feature extraction of the ground remote sensing image to obtain spectral features according to the local maximum point sequence includes: projecting the local maximum point sequence into the eigenspace and the penalty space to obtain the eigengraph and the penalty graph respectively based on the constructed intra-class neighbor samples; and characterizing the spectral structure of the eigengraph and the penalty graph based on the regional median filtering to extract the features of the ground remote sensing image to obtain the spectral features. To this end, the above scheme has the following beneficial technical effects: by projecting the local maximum point sequence into the eigenspace and the penalty space, the spectral differences of different categories or types of objects in the image can be highlighted. This helps subsequent classification or recognition tasks to more accurately identify different ground objects or regions. Using the intra-class neighbor samples to construct the projection space makes the feature extraction process more robust to noise and outliers. This method takes into account the similarity between samples of the same type, thereby improving the stability and reliability of feature extraction. Characterizing the spectral structure of the eigengraph and the penalty graph by regional median filtering can effectively retain the local and global spectral structure information in the image, which is crucial for analyzing complex scenes and distribution patterns of objects in ground remote sensing images. Projecting the sequence of local maximum points into a low-dimensional space can significantly reduce the amount of data for subsequent processing, thereby reducing computational complexity. This is of great significance for processing large-scale ground remote sensing image data, and can improve the efficiency of feature extraction and subsequent analysis. By constructing eigengraphs and penalty graphs, the distribution of different categories or types of ground objects in the spectral feature space can be intuitively displayed, which helps researchers better understand the spectral characteristics of ground remote sensing images and provides strong support for subsequent analysis and interpretation. This method does not depend on a specific remote sensing image data source or ground object type and has strong adaptability. It can be applied to different types of ground remote sensing image data, including images with different resolutions, different band combinations, and different ground object coverage types.

[0064] This application provides an exemplary code as follows:

[0065]

[0066]

[0067] Optionally, the spectral structure characterization of the eigenmap and the penalty map based on regional median filtering to extract features from the ground remote sensing image to obtain spectral features includes: characterizing the spectral structure of the eigenmap and the penalty map based on regional median filtering to obtain spectral fluctuation features; performing dual-tree complex wavelet transform on the spectral fluctuation features to perform multi-scale decomposition on the spectral fluctuation features and obtain multi-scale multi-directional features; performing dynamic threshold density clustering segmentation on the ground remote sensing image based on the multi-scale multi-directional features to extract spectral features therefrom. To this end, the above scheme has the following beneficial technical effects: Enhanced spectral structure detail capture: The spectral structure characterization of the eigenmap and the penalty map by regional median filtering can effectively eliminate noise and outliers, while retaining the local structural information of the spectral features, which helps to more accurately capture the spectral fluctuation features in the ground remote sensing image, and provide more refined information for subsequent feature extraction and analysis. Multi-scale multi-directional feature extraction: Dual-Tree Complex Wavelet Transform (DT-CWT) on the spectral fluctuation features can achieve multi-scale multi-directional feature extraction. DT-CWT can not only provide multi-scale analysis similar to traditional wavelet transform, but also capture the directional characteristics of the signal through the phase information of the complex wavelet coefficients, which helps to extract richer and more comprehensive spectral features from ground remote sensing images, including spectral change information at different scales and directions. Dynamic threshold density clustering segmentation: Dynamic threshold density clustering segmentation of ground remote sensing images based on multi-scale and multi-directional features can achieve adaptive segmentation of different spectral feature areas. This method can dynamically adjust the clustering threshold and density parameters according to the distribution and complexity of the spectral features in the image, so as to obtain more accurate and reasonable segmentation results, which helps to extract more accurate and effective spectral features from ground remote sensing images, and provide strong support for subsequent classification, recognition and other tasks. Improve the robustness and adaptability of feature extraction: The entire process combines multiple technologies such as regional median filtering, even tree complex wavelet transform and dynamic threshold density clustering segmentation, each of which has different advantages and scope of application. Using them in combination can complement each other and work together to improve the robustness and adaptability of feature extraction, enabling this method to cope with ground remote sensing image data of different complexity and noise levels and extract more accurate and effective spectral features.

[0068] This application provides an exemplary code as follows:

[0069]

[0070]

[0071] Optionally, extracting features from the ground remote sensing image to obtain texture features includes:

[0072] Pixels are extracted from the ground remote sensing image, and the grayscale correlation between the pixels is calculated; grayscale correlation texture feature vectors are generated according to the grayscale correlation between the pixels; the ground remote sensing image is segmented to obtain a number of image blocks, and the difference distribution of the internal pixel values ​​of each image block is calculated; based on the constructed grayscale co-occurrence matrix, the difference distribution of the internal pixel values ​​of all image blocks is calculated to generate a differential correlation texture feature vector; the grayscale correlation texture feature vector and the differential correlation texture feature vector are fused to obtain a joint texture feature vector; according to the joint texture feature vector, feature extraction is performed on the ground remote sensing image to obtain texture features. To this end, the above scheme has the following beneficial technical effects: by calculating the grayscale correlation between pixels and the difference distribution of the internal pixel values ​​of the image blocks, rich texture information can be extracted. This information is crucial for subsequent tasks such as ground object recognition, classification and scene understanding. The fusion of grayscale correlation texture feature vectors and differential correlation texture feature vectors forms a joint texture feature vector. This fusion of multiple features can more comprehensively characterize the texture characteristics of the image, thereby improving the accuracy of image recognition and classification. Since texture features are usually relatively independent of attributes such as shape, size, and color of objects in the image, they are robust to changes in conditions such as lighting, shadows, and occlusions, making the analysis results of remote sensing images more reliable.

[0073] This application provides an exemplary code as follows:

[0074]

[0075]

[0076] Optionally, the pixel extraction of the ground remote sensing image and the calculation of the grayscale correlation between the pixels include: based on the set grayscale threshold, the pixel extraction of the ground remote sensing image; based on the covariance of the grayscale difference between each pixel and its neighboring pixels, the grayscale correlation between the pixels is calculated. To this end, the above scheme has the following beneficial technical effects: by setting the grayscale threshold for pixel extraction, the noise interference in the image can be eliminated or weakened. This threshold operation helps to highlight the useful information in the image while suppressing the influence of noise on the subsequent grayscale correlation calculation. The selection of the grayscale threshold can be adjusted according to the characteristics of the image and the analysis requirements, so as to retain the important details in the image, which helps to more accurately identify the ground objects in the subsequent feature extraction and classification tasks. By calculating the covariance of the grayscale difference between each pixel and its neighboring pixels, the local texture features in the image can be captured. This covariance calculation method based on grayscale difference is sensitive to texture details and helps to extract texture features useful for classification and recognition. Since grayscale correlation is calculated based on the grayscale difference between pixels, this method is robust to the influence of lighting changes, shadows and other conditions, so that the extracted texture features can remain relatively stable under different conditions. The extracted grayscale correlation features can be fused with other types of features (such as shape features, color features, etc.) to form a more comprehensive and integrated feature vector. This fusion method helps to improve the accuracy and robustness of classification and recognition.

[0077] This application provides an exemplary code as follows:

[0078]

[0079] Optionally, the grayscale correlation texture feature vector is generated according to the grayscale correlation between the pixels, including: calculating the grayscale second-order moment, entropy and contrast features of the pixels according to the grayscale correlation between the pixels; fusing the grayscale second-order moment, entropy and contrast features of the pixels to obtain the grayscale correlation texture feature vector. To this end, the above scheme has the following beneficial technical effects: by calculating multiple texture features such as grayscale second-order moment, entropy and contrast, texture information of different aspects in the image can be captured, and these features together form a rich feature set, which can more comprehensively describe the texture characteristics of the image. Features such as grayscale second-order moment, entropy and contrast have certain robustness in describing image texture. They are less sensitive to factors such as illumination changes and noise interference, and can maintain relatively stable performance under different image conditions. These texture features can effectively distinguish different types of objects or targets. For example, the grayscale second-order moment reflects the uniformity of the grayscale distribution of the image, entropy describes the complexity of the grayscale distribution of the image, and contrast highlights the difference in grayscale values ​​in the image. These features will show different values ​​in different texture modes, thereby achieving effective distinction between different textures. Features such as grayscale second-order moment, entropy, and contrast have clear physical meanings and are easy to understand and explain, making texture analysis based on these features more intuitive and reliable, and helpful for in-depth analysis and understanding of texture information in images. By fusing features such as grayscale second-order moment, entropy, and contrast, a comprehensive grayscale-related texture feature vector can be obtained. This feature vector not only contains information about various texture features, but also can further improve the stability and distinguishing ability of the features through fusion, and can make full use of the advantages of various texture features, making the final texture analysis more accurate and reliable.

[0080] This application provides an exemplary code as follows:

[0081]

[0082] Optionally, performing image segmentation on the ground remote sensing image to obtain a plurality of image blocks, and calculating the difference distribution of internal pixel values ​​for each image block, includes: performing image segmentation on the ground remote sensing image to obtain a plurality of image blocks based on a region growing multi-scale segmentation algorithm; merging the plurality of image blocks based on regional similarity to obtain image blocks to be processed, and calculating the difference distribution of internal pixel values ​​for each image block to be processed.

[0083] Optionally, the method of calculating the difference distribution of internal pixel values ​​of all image blocks based on the constructed gray level co-occurrence matrix to generate a differential correlation texture feature vector includes: calculating the difference distribution of internal pixel values ​​of all image blocks based on the constructed gray level co-occurrence matrix, calculating the distance adjacency distribution characteristics of pixel values ​​in space and the angle adjacency distribution characteristics of pixel values; fusing the distance adjacency distribution characteristics of pixel values ​​and the angle adjacency distribution characteristics of pixel values ​​to obtain a differential correlation texture feature vector. To this end, the above scheme has the following beneficial technical effects: the region growing multi-scale segmentation algorithm can segment images at different scales, thereby capturing the features of objects at different scales. This is very important for remote sensing images, because the size and shape of objects may vary greatly, and multi-scale segmentation can better adapt to such changes. Through image segmentation, the entire image can be divided into several smaller image blocks. Then, subsequent processing can be performed independently for each image block, such as calculating the difference distribution of pixel values, thereby significantly improving the computational efficiency. Image segmentation can limit noise and interference factors to a smaller area, thereby reducing the impact on the overall analysis. When calculating the difference distribution of pixel values ​​in the subsequent process, since only the pixels inside each image block are concerned, the features of the objects can be better highlighted and the interference of noise can be reduced. The merging process based on regional similarity can further optimize the division of image blocks, so that the merged image blocks to be processed are more consistent with the actual distribution of objects, which helps to enhance the accuracy of subsequent feature extraction. For example, when calculating the difference distribution of pixel values, the texture, shape and other information of the objects can be more accurately reflected. By calculating the difference distribution of internal pixel values ​​for each image block to be processed, rich feature information such as texture and structure can be extracted. These feature information are crucial for subsequent object classification and recognition tasks, and help to improve the accuracy of classification and recognition.

[0084] This application provides an exemplary code as follows:

[0085]

[0086] In the above code, the slic function is a multi-scale segmentation algorithm in the scikit-image library, which is based on the Simple Linear Iterative Clustering (SLIC) method. The merged_segments variable should contain the labels of the image patches after merging based on region similarity.

[0087] Optionally, the grayscale-related texture feature vector and the differential-related texture feature vector are fused to obtain a joint texture feature vector, including: calculating the entropy of the grayscale-related texture feature vector and the differential-related texture feature vector; determining the respective fusion weights according to the entropies corresponding to the grayscale-related texture feature vector and the differential-related texture feature vector; and weightedly fusing the grayscale-related texture feature vector and the differential-related texture feature vector based on the respective fusion weights of the grayscale-related texture feature vector and the differential-related texture feature vector to obtain a joint texture feature vector. To this end, the above scheme has the following beneficial technical effects: the grayscale-related texture feature vector and the differential-related texture feature vector respectively describe the texture information of the image from different angles. The grayscale-related texture feature focuses on the grayscale correlation between pixels, while the differential-related texture feature focuses on the change of pixel values. By fusing these two feature vectors, their respective advantages can be comprehensively utilized to obtain more comprehensive and rich texture information. By using the entropy of the feature vector to determine the fusion weight, an adaptive fusion method can be implemented. Entropy is an indicator to measure the amount of information. By calculating the entropy of grayscale-related texture feature vectors and differential-related texture feature vectors, the amount of information they each contain can be evaluated. Determining the fusion weight according to the size of the entropy can make the feature vector with more information occupy a larger weight in the fusion process, thereby improving the effectiveness and accuracy of the joint texture feature vector. Due to the combination of two different types of texture features, the joint texture feature vector has better robustness to interference factors such as noise and illumination changes in the image. Even if one of the feature vectors is distorted by interference, the other feature vector can still provide useful information, thereby maintaining the stability and reliability of the joint texture feature vector. Improve classification and recognition performance: By fusing the grayscale-related texture feature vector and the differential-related texture feature vector, the joint texture feature vector obtained can more accurately describe the texture characteristics of the image, which helps to improve the performance of tasks such as image classification and target recognition, and enables the algorithm to better adapt to complex and changeable image data.

[0088] This application provides an exemplary code as follows:

[0089]

[0090] Optionally, the method of extracting features from the ground remote sensing image to obtain texture features according to the joint texture feature vector comprises: performing a non-subsampled wavelet transform on the ground remote sensing image to obtain a feature coefficient matrix based on the constructed wavelet basis according to the joint texture feature vector, and obtaining texture features based on the variance of the feature coefficient matrix. To this end, the above scheme has the following beneficial technical effects: the non-subsampled wavelet transform provides a multi-scale analysis of the image, which means that the texture information of the image can be captured at different frequencies and resolutions at the same time. This is particularly important for ground remote sensing images because different texture features may be more obvious at different scales. A key feature of the non-subsampled wavelet transform is its translation invariance. This means that a slight change in the position of the texture features in the image will not significantly affect the result of the transformation, thereby improving the robustness of feature extraction. Compared with the traditional wavelet transform, the non-subsampled wavelet transform does not reduce the size of the image during the transformation process, thereby reducing information loss, helping to retain subtle texture features in the image and improving the accuracy of subsequent analysis. The variance of the feature coefficient matrix is ​​a simple but effective measure for quantifying the intensity or degree of variation of texture features at different scales. This measure is easy to calculate and can provide useful information about texture complexity and consistency. By combining the joint texture feature vector and the non-subsampled wavelet transform, richer and more diverse texture features can be extracted. These features not only contain the grayscale information and differential information in the original image, but also contain the texture variation information at multiple scales, providing more powerful support for subsequent image analysis or classification. By extracting these multi-scale and multi-directional texture features, the accuracy of ground remote sensing image classification can be significantly improved. These features can better describe the local and global texture patterns in the image, thereby helping to distinguish different types of objects or scenes.

[0091] This application provides an exemplary code as follows:

[0092]

[0093] Optionally, the feature extraction of the ground remote sensing image to obtain context features includes: based on the training completed embedded hybrid attention block, feature extraction of shallow image features and deep image features for the ground remote sensing image; based on the shallow image features and deep image features, feature context feature characterization of the ground remote sensing image to obtain context features. To this end, the above scheme has the following beneficial technical effects: by extracting shallow image features and deep image features at the same time, image information at different levels can be fused. Shallow features usually contain local details and edge information of the image, while deep features focus more on the global structure and semantic information of the image. This fusion of multi-level features helps to more comprehensively describe the context information of the image. The design of the embedded hybrid attention block enables the model to focus on key areas in the image, that is, those parts that are more important for context feature characterization. This attention mechanism can enhance the model's ability to capture important information in the image and improve the accuracy and efficiency of feature extraction. Based on the extracted shallow image features and deep image features, the ground remote sensing image is characterized in context, and more abundant and comprehensive context information can be obtained. This contextual feature can not only reflect the local texture and structural information in the image, but also describe the spatial relationship and semantic connection between different objects in the image, providing strong support for subsequent image analysis and understanding. By extracting features through the trained embedded hybrid attention block, the model can learn a more robust and generalized feature representation. This feature representation is not only applicable to training data, but can also be well generalized to unseen test data, improving the performance of the model in practical applications.

[0094] This application provides an exemplary code as follows:

[0095]

[0096]

[0097] Optionally, extracting features from the ground remote sensing image to obtain shape features includes:

[0098] The ground remote sensing image is subjected to edge detection to obtain an edge image; the edge image is subjected to corner point detection, and the detected corner points are merged based on the Euclidean distance to obtain a distance map; the distance map is calculated with an invariant distance to obtain a shape feature vector; and the ground remote sensing image is subjected to feature extraction based on the shape feature vector to obtain a shape feature. To this end, the above scheme has the following beneficial technical effects: by edge detection, the shape boundary in the image can be highlighted, redundant information can be removed, and subsequent shape feature extraction can be more focused on the shape of the image itself. Corner points are generally considered to be one of the most important feature points in an image because they represent the places where the shape changes most dramatically in the image. By detecting corner points, the key features of the image shape can be captured. By merging corner points based on the Euclidean distance, redundant corner points generated by noise or over-detection can be further reduced, making the shape features more concise and accurate. Invariant distance is a shape-based geometric feature descriptor that is invariant to image transformations such as translation, rotation and scaling. By calculating the invariant distance of the distance map, a feature vector that is robust to shape changes can be obtained. Since shape features can capture the basic shape information of objects in the image, image recognition based on shape features usually has high accuracy. Especially in ground remote sensing images, many ground objects have unique shape features, and using these features for recognition can greatly improve the accuracy and efficiency of recognition. Ground remote sensing images usually contain a variety of complex ground objects, such as buildings, roads, farmland, etc. These objects may have large differences in shape, but by extracting shape features, they can be effectively distinguished and accurate recognition of different objects in complex scenes can be achieved.

[0099] This application provides an exemplary code as follows:

[0100]

[0101]

[0102] Optionally, performing edge detection on the ground remote sensing image to obtain an edge image includes:

[0103] The pixel value gradient and direction of the ground remote sensing image are calculated, and the calculated pixel value gradient and direction are non-maximum suppressed to obtain an edge image. To this end, the above scheme has the following beneficial technical effects: the calculation of the pixel value gradient can accurately indicate the position where the brightness changes most dramatically in the image, that is, the edge. These positions are often the key information of the boundary of the ground object, which is crucial for subsequent image analysis and understanding. Non-maximum suppression is an important step. It compares the size of the current pixel point with its neighboring pixel points in the gradient direction, retains only the local maximum value in the gradient direction, thereby eliminating those non-edge points, so that pseudo edges caused by noise and texture can be removed, making the detected edges clearer and more accurate. When calculating the pixel value gradient, changes in multiple directions are usually considered. This allows the edge detection algorithm to adapt to edges in different directions, whether they are horizontal, vertical or oblique. This is very useful for edges of various shapes and directions that may appear in ground remote sensing images. The gradient-based edge detection method is relatively robust to changes in illumination. Because the gradient calculates the rate of change of the pixel value, rather than the absolute value of the pixel value, even if the illumination conditions change, as long as the edges of the ground object still exist, the edge detection algorithm can detect them. Edge detection methods based on gradient and non-maximum suppression are relatively efficient in computation. They usually only need to traverse the image once to obtain the edge image, and the algorithm complexity is low, which is suitable for processing large-scale ground remote sensing images.

[0104] This application provides an exemplary code as follows:

[0105]

[0106]

[0107] Use a Gaussian filter to smooth the image to reduce the impact of image noise on edge detection. Calculate the gradient magnitude and direction of the image. Apply non-maximum suppression to retain only the local maximum in the gradient direction. Use a dual threshold algorithm for edge detection and connection. In the above code, threshold1 and threshold2 are two threshold parameters in Canny edge detection. A smaller threshold is used to capture weak edges, while a larger threshold is used to capture strong edges. Edges between these two thresholds are retained based on their connectivity with the strong edge.

[0108] Optionally, the corner point detection of the edge image and merging the detected corner points based on the Euclidean distance to obtain a distance map include: determining the edge contour on the edge image; calculating the distance between the pixel point on the edge contour and a chord formed by connecting the head and tail endpoints, and taking the point with the longest distance as the corner point; and merging the detected corner points based on the Euclidean distance based on the image edge detection operator constructed on the Euclidean distance map to obtain the distance map. To this end, the above scheme has the following beneficial technical effects: by calculating the distance between the pixel point on the edge contour and the chord formed by connecting the head and tail endpoints, and taking the point with the longest distance as the corner point, this method can accurately identify the corner points in the image. Corner points are usually the points with the most dramatic shape changes in the image, which are of great significance for subsequent image analysis and recognition. Merging corner points based on the Euclidean distance can effectively reduce redundant corner points caused by noise or over-detection. By merging similar corner points, a more concise and accurate set of corner points can be obtained, thereby improving the efficiency and accuracy of subsequent processing. The method has certain robustness to changes in illumination, rotation, scaling, etc. in the image. Since corner detection is based on edge contours, it is not affected by changes in image brightness; the merging process based on Euclidean distance can handle small displacements and shape changes of corners. By merging corners to obtain a distance map, complex edge contours can be simplified into a series of corners and their relative positional relationships. This simplified shape representation is conducive to subsequent image matching, recognition, and classification tasks. Although the corner detection and merging process requires a certain amount of computation, this method has higher computational efficiency than other complex image processing algorithms. It can quickly process large-scale image data and extract useful shape features. Ground remote sensing images usually contain a variety of complex ground objects, which may have large differences in shape. Through the corner detection and merging method based on edge contours and Euclidean distance, the key feature points of these complex shapes can be effectively extracted, thereby achieving accurate recognition and analysis of different objects.

[0109] This application provides an exemplary code as follows:

[0110]

[0111]

[0112] Optionally, the calculation of the invariant distance for the distance map to obtain the shape feature vector includes: calculating the centroid of the distance map and determining the distance from each pixel point on the contour of the distance map to the centroid; determining the invariant features of the shape on the distance map according to the distance from each pixel point to the centroid; generating the shape feature vector according to the invariant features of the shape on the distance map. To this end, the above scheme has the following beneficial technical effects: the invariant distance (such as geometric moment, Zernike moment, etc.) has scale invariance, which means that even if the size of the shape in the image changes, the calculated invariant distance can still remain unchanged. This is very useful for processing shapes in images of different resolutions or scales. Invariant distances, such as Zernike moments, also have rotation invariance. This means that no matter how the shape is rotated in the image, its invariant distance can remain unchanged. This is particularly useful for processing shapes in images taken at different angles. Invariant distances have a certain robustness to noise and slight deformation. This means that even if the shape is disturbed or deformed to a certain extent, its invariant distance can still remain relatively stable, so that the shape can be accurately represented. The shape feature vector is a compact, low-dimensional representation that contains the main information of the shape. This makes it more efficient to store and transfer shape data, and also facilitates its use in subsequent analysis and recognition tasks. Shape feature vectors can be used directly to compare the similarities between different shapes. By calculating the distance or similarity measure between two shape feature vectors, it is possible to quickly determine whether two shapes are similar or identical.

[0113] This application provides an exemplary code as follows:

[0114]

[0115]

[0116] Optionally, the method of extracting features from the ground remote sensing image to obtain shape features based on the shape feature vector includes: generating a shape matrix based on the shape feature vector based on coordinate grating resampling; and performing Fourier transform on the shape matrix to obtain a shape descriptor as the shape feature. To this end, the above scheme has the following beneficial technical effects: the shape matrix can be accurately generated based on the shape feature vector by a method based on coordinate grating resampling. This method can maintain the integrity and accuracy of shape information and provide a reliable basis for subsequent feature extraction. Fourier transform has rotation and translation invariance, which means that even if the shape in the image is rotated or translated, the shape descriptor obtained by Fourier transform can still remain unchanged. This is particularly important for shape feature extraction in ground remote sensing images, because the target objects in remote sensing images often undergo rotations and position changes at different angles. Fourier transform can quickly calculate shape descriptors, which helps to improve the computational efficiency of the entire feature extraction process and makes it possible to process large-scale ground remote sensing image data. The shape descriptor obtained by Fourier transform contains rich shape information, such as the size, direction, contour, etc. of the shape. This information can fully describe the shape characteristics of the target in the ground remote sensing image, providing strong support for subsequent target recognition, classification and other tasks. This method has a certain robustness to noise and slight deformation. Even if there is a certain degree of noise in the image or the target shape is slightly deformed, the shape descriptor obtained by Fourier transform can still remain relatively stable, thus ensuring the accuracy of feature extraction.

[0117] This application provides an exemplary code as follows:

[0118]

[0119]

[0120] Optionally, the spectral feature and the texture feature are fused to obtain a first fused feature, including: performing convolution smoothing on the spectral feature and the texture feature to project them into a standard normal variable transformation space, and performing continuous projection in the standard normal variable transformation space to obtain the first fused feature. To this end, the above scheme has the following beneficial technical effects: by projecting the spectral feature and the texture feature into the standard normal variable transformation space, it can be ensured that features of different sources and scales have the same statistical distribution. This helps to eliminate the scale differences and biases between features during the fusion process, making the fused features more unified and comparable. The standard normal variable transformation can usually reduce the correlation between features, which helps to reduce the risk of overfitting in subsequent machine learning or classification tasks, because features with high correlation may contain redundant information, and removing this redundant information can improve the generalization ability of the model. Convolution smoothing can smooth and enhance features, remove noise and unnecessary details, while retaining important shape and structure information, which helps to improve the robustness and stability of features, making the fused features more suitable for various application scenarios. Continuous projection in the standard normal variable transformation space can achieve efficient feature fusion. Through continuous projection, multiple features can be gradually fused into a unified feature vector while maintaining the important information in the original features. This method does not require complex feature selection or combination strategies and can automatically find the optimal feature fusion method. Since the fused features contain both spectral and texture information, they can more comprehensively describe the characteristics of ground remote sensing images. This helps to improve the performance of the classifier, especially when dealing with complex land object classification problems, the fused features can provide more discriminant information, making the classification results more accurate and reliable.

[0121] This application provides an exemplary code as follows:

[0122]

[0123]

[0124] Optionally, the fusion of the context features and shape features to obtain the second fused features includes: performing convolution smoothing on the context features and shape features to project them into a standard normal variable transformation space, and performing continuous projection in the standard normal variable transformation space to obtain the second fused features. To this end, the above scheme has the following beneficial technical effects: through standard normal normalization, context features and shape features of different scales and sources are converted to the same scale so that they have the same weight when fused. This helps to reduce the deviation caused by the difference in feature scales and improve the accuracy of fused features. Convolution smoothing can remove noise and redundant information in features while retaining important shape and structural information, which helps to improve the robustness of fused features and make them more stable and reliable. Continuous projection in the standard normal variable transformation space can efficiently fuse multiple features into a unified feature vector. This fusion method does not require complex feature selection or combination strategies and can automatically find the optimal feature fusion method. Since the fused features contain both contextual information and shape information, they can more comprehensively describe the characteristics of the target object, which helps to improve the performance of the classifier or recognition system. Especially when dealing with complex scenes or target objects, the fused features can provide more discriminant information, making the classification or recognition results more accurate.

[0125] This application provides an exemplary code as follows:

[0126]

[0127] Optionally, generating a vector terrain network according to the first fused feature and the second fused feature includes: rendering the first fused feature and the second fused feature into a texture pyramid to obtain a vector elevation, and drawing a vector terrain network based on the vector elevation. To this end, the above scheme has the following beneficial technical effects: by rendering the fused feature into the texture pyramid, the multi-level details of the terrain can be captured. The texture pyramid can efficiently represent the features of the terrain at different scales by storing terrain data at different resolutions, thereby achieving high-precision terrain representation. The data structure of the texture pyramid allows only the necessary level of detail to be loaded when drawing the terrain, thereby reducing memory usage and computational burden. This makes it more efficient to draw a vector terrain network on large-scale terrain data. By fusing spectral features, texture features, context features, and shape features, richer terrain information can be obtained. Rendering this information into the texture pyramid can produce a more realistic and detailed terrain visualization effect, thereby improving the user experience. The vector terrain network allows users to perform various interactive operations, such as zooming, rotating, and panning. Since the texture pyramid supports fast data access and updating, these interactive operations can be performed at real-time or near real-time speeds, improving the responsiveness of the system.

[0128] This application provides an exemplary code as follows:

[0129]

[0130] Optionally, extracting the ground object on the remote sensing image according to the vector terrain network includes: determining the halo on the remote sensing image based on the vector terrain network based on a color key set; generating a three-dimensional terrain vector and a 2D space vector based on the halo; and projecting the three-dimensional terrain vector and the 2D space vector into a terrain grid to extract the ground object on the remote sensing image. To this end, the above scheme has the following beneficial technical effects: using the vector terrain network as a reference, the boundary and position of the ground object on the remote sensing image can be accurately determined, and this method overcomes the problems of edge blur and insufficient precision that may be caused by the traditional pixel-based extraction method. By determining the halo through the color key set, the specific color area related to the terrain in the remote sensing image can be quickly and accurately identified, thereby improving the processing efficiency and accuracy. The generated three-dimensional terrain vector provides stereo information of the terrain, while the 2D space vector retains the shape and position of the ground object on the plane. The combination of the two makes the extracted ground object not only highly accurate, but also contains rich spatial information. Projecting the 3D terrain vectors and 2D space vectors into the terrain grid can ensure that the extracted ground objects are consistent with the real terrain. The terrain grid-based extraction method can better handle terrain changes and the complexity of ground objects.

[0131] This application provides an exemplary code as follows:

[0132]

[0133]

[0134] The scope of the invention involved in the application is not limited to the technical solutions formed by the specific combination of the above technical features, but also includes other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for extracting ground objects based on remote sensing images. It is characterized in that include: Acquisition of ground remote sensing images; Extracting features from the ground remote sensing image to obtain spectral features, texture features, context features, and shape features; Fusing the spectral feature and the texture feature to obtain a first fused feature; Fusing the context feature and the shape feature to obtain a second fused feature; generating a vector terrain network according to the first fused features and the second fused features; extracting the ground objects on the remote sensing image according to the vector terrain network; The step of extracting features from the ground remote sensing image to obtain spectral features includes: Calculating pixel density distribution on the ground remote sensing image; Determining a plurality of local maximum points of pixels on the ground remote sensing image according to the pixel density distribution; Based on the positions of the pixels, the plurality of local maximum points of the pixels are serialized to obtain a local maximum point sequence; According to the local maximum point sequence, feature extraction is performed on the ground remote sensing image to obtain spectral features; Wherein, the calculating pixel density distribution on the ground remote sensing image comprises: For any pixel point on the ground remote sensing image, obtain its pixel value, and determine the pixel values ​​of all pixel points in the neighborhood constructed for the any pixel point; Based on the formula Calculate the pixel density with any pixel point as reference, Represents the pixel value of the i-th pixel on the ground remote sensing image, is the pixel value of any pixel in the neighborhood of the i-th pixel, is the number of pixels considered in the neighborhood, is the bandwidth parameter, used to determine the size of the neighborhood, is the kernel function, is the pixel density of the i-th pixel; Calculating the pixel density gradient distribution of any pixel point on the ground remote sensing image based on the pixel density of all pixels as reference; The pixel density gradient distribution of any pixel point on the ground remote sensing image is calculated based on the pixel density of all pixel points referenced by the following formula, including: , in, is the total number of pixels in the neighborhood, is the weight, yes Kernel function, is the bandwidth parameter used for smoothing, and are the pixel densities of the j-th pixel and the i-th pixel, respectively. is the pixel density gradient distribution of the i-th pixel; Wherein, determining a plurality of local maximum points of pixels on the ground remote sensing image according to the pixel density distribution includes: Based on the set iteration step, according to the formula The pixel value of the i-th pixel Iterate, where Indicates the current iteration number, represents the number of next iteration steps, Indicates the pixel value corresponding to the current iteration step, Indicates the corresponding pixel value at the next iteration step, is a positive step size parameter, Indicates the pixel density corresponding to the pixel value at the current iteration step; Calculate the pixel density update value corresponding to the corresponding pixel value at the next iteration step , if satisfied , then the pixel density of the pixel point is a local maximum point of a pixel, is the set threshold.

2. The method according to claim 1, It is characterized in that The step of serializing the plurality of local maximum value points of pixels based on the positions of the pixels to obtain a local maximum value point sequence includes: Count the feature intensity of each pixel's local maximum point; Based on the positions of the pixels, the characteristic intensities of the local maximum points of the pixels are serialized to obtain a local maximum point sequence; The statistical feature strength of each pixel local maximum point includes: Calculate the ratio of the pixel density of each pixel's local maximum point to the mean of the pixel density in its neighborhood; Calculate the variance of the pixel density of each pixel's local maximum point and the pixel density in its neighborhood; Based on the formula , weighted operation is performed on the ratio and the variance to calculate the characteristic intensity of the local maximum point of each pixel, wherein, is the characteristic intensity of the local maximum point of each pixel, is the ratio of each pixel’s local maximum point to the mean pixel density in its neighborhood. is the variance of pixel density in the neighborhood, and is a weight parameter used to balance the contribution of ratio and variance to feature strength.

3. The method according to claim 2, It is characterized in that The method of serializing the characteristic intensities of the plurality of local maximum value points of the pixels based on the positions of the pixels to obtain a local maximum value point sequence includes: Construct a feature vector based on the position of the pixel and the local maximum point of the pixel corresponding to the position; Collecting the eigenvectors of all local maximum points into a matrix, wherein each row represents a vector of a local maximum point, to obtain a principal component analysis matrix, wherein the principal component analysis matrix has zero mean and unit variance; The eigenvalues ​​of the principal component analysis matrix are calculated to serialize the plurality of local maximum points of pixels based on the eigenvalues ​​to obtain a local maximum point sequence.

4. The method according to claim 3, It is characterized in that The step of extracting features from the ground remote sensing image according to the local maximum point sequence to obtain spectral features includes: Based on the constructed in-class nearest neighbor samples, the local maximum point sequence is projected into the eigenspace and the penalty space to obtain an eigengraph and a penalty graph respectively; Based on regional median filtering, the spectral structure of the intrinsic image and the penalty image is characterized to extract features of the ground remote sensing image to obtain spectral features; The method of performing spectral structure characterization on the eigenimage and the penalty image based on regional median filtering to extract features of the ground remote sensing image to obtain spectral features includes: Based on regional median filtering, the spectral structure of the intrinsic graph and the penalty graph is characterized to obtain spectral fluctuation characteristics; Performing even-tree complex wavelet transform on the spectral fluctuation feature to perform multi-scale decomposition on the spectral fluctuation feature and obtain multi-scale and multi-directional features; Based on the multi-scale and multi-directional features, dynamic threshold density clustering segmentation is performed on the ground remote sensing image to extract spectral features therefrom; The step of extracting features from the ground remote sensing image to obtain texture features includes: Extracting pixels from the ground remote sensing image and calculating the grayscale correlation between the pixels; According to the grayscale correlation between the pixels, a grayscale correlation texture feature vector is generated; Performing image segmentation on the ground remote sensing image to obtain a plurality of image blocks, and calculating the difference distribution of pixel values ​​within each image block; Based on the constructed gray-level co-occurrence matrix, the difference distribution of the internal pixel values ​​of all image blocks is calculated to generate the differential correlation texture feature vector; fusing the grayscale-related texture feature vector and the difference-related texture feature vector to obtain a joint texture feature vector; According to the joint texture feature vector, feature extraction is performed on the ground remote sensing image to obtain texture features; The step of extracting features from the ground remote sensing image to obtain context features includes: Based on the training completed embedded hybrid attention block, feature extraction of shallow image features and deep image features is performed on the ground remote sensing image; Based on the shallow image features and the deep image features, characterizing the ground remote sensing image by feature context to obtain context features; The step of extracting features from the ground remote sensing image to obtain shape features includes: Performing edge detection on the ground remote sensing image to obtain an edge image; Performing corner point detection on the edge image, and merging the detected corner points based on Euclidean distance to obtain a distance map; Calculating an invariant distance on the distance map to obtain a shape feature vector; According to the shape feature vector, feature extraction is performed on the ground remote sensing image to obtain shape features; The step of performing edge detection on the ground remote sensing image to obtain an edge image includes: The pixel value gradient and direction of the ground remote sensing image are calculated, and non-maximum suppression is performed on the calculated pixel value gradient and direction to obtain an edge image.

5. The method according to claim 4, It is characterized in that The performing corner point detection on the edge image and merging the detected corner points based on the Euclidean distance to obtain a distance map includes: determining an edge contour on the edge image; For the distance between a pixel point on the edge contour and a chord formed by connecting the first and last endpoints, the point with the longest distance is taken as a corner point; Based on the constructed Euclidean distance map, the image edge detection operator merges the detected corner points based on the Euclidean distance to obtain a distance map.

6. The method according to claim 5, It is characterized in that The step of calculating the invariant distance on the distance map to obtain a shape feature vector includes: Calculating the centroid of the distance map, and determining the distance from each pixel point on the contour of the distance map to the centroid; Determining invariant features of shapes on the distance map according to the distances from each pixel to the centroid; The shape feature vector is generated according to the invariant features of the shape on the distance map.

7. The method according to claim 6, It is characterized in that The step of extracting features from the ground remote sensing image according to the shape feature vector to obtain shape features includes: Based on coordinate raster resampling, a shape matrix is ​​generated according to the shape feature vector; Performing Fourier transform on the shape matrix to obtain a shape descriptor as the shape feature.

8. The method according to claim 7, It is characterized in that The fusing of the spectral feature and the texture feature to obtain the first fused feature comprises: performing convolution smoothing processing on the spectral feature and the texture feature to project them into a standard normal variable transformation space, and performing continuous projection in the standard normal variable transformation space to obtain the first fused feature; The fusing of the context features and shape features to obtain a second fused feature includes: performing convolution smoothing on the context features and shape features to project them into a standard normal variable transformation space, and performing continuous projection in the standard normal variable transformation space to obtain a second fused feature.

Citation Information

Patent Citations

  • Intelligent classification method for high-resolution remote sensing image artificial structure building scene

    CN117876878A

  • Unsupervised Latent Low-Rank Projection Learning Method for Feature Extraction of Hyperspectral Images

    US20230114877A1