Remote sensing image multi-scale object-oriented parallel segmentation method

Abstract

The invention discloses a remote sensing image multi-scale object-oriented parallel segmentation method, and relates to the technical field of remote sensing image segmentation. The remote sensing image multi-scale object-oriented parallel segmentation method comprises the following steps: obtaining image data and carrying out slicing processing to obtain a plurality of slice blocks; then performing complexity analysis on each slice block, calculating task load indexes and sorting the task load indexes; according to the number of CPU cores and a preset rule, tasks are distributed, and image segmentation parallel processing is carried out; according to the method, the remote sensing image data is segmented into the multiple slice blocks, parallel task allocation is carried out, the processing time can be effectively shortened, priority ranking is carried out on the slice blocks according to the complexity of the slice blocks, and therefore it can be ensured that computing resources are allocated in a balanced mode according to task loads; the situation that some complex tasks drag down the overall speed is avoided, the segmentation process of each slice block is processed in parallel, and the execution speed of large-scale image segmentation can be remarkably increased.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image segmentation, in particular to a remote sensing image multi-scale object-oriented parallel segmentation method. Background Art

[0002] Object-oriented image segmentation and classification can better maintain the edges, textures and spectral characteristics of homogeneous objects by segmenting pixels into homogeneous regions and distinguishing different land types. It overcomes the problems of image fragmentation, boundary confusion and high redundancy after traditional pixel-by-pixel classification. It is highly favored in the extraction of medical and remote sensing image information. However, due to the large volume of high-resolution remote sensing image data, the object-oriented cutting method has a complex calculation process and a slow calculation speed.

[0003] Prior art, such as the invention patent application with announcement number: CN116128910A, discloses a multi-scale segmentation method for remote sensing images based on spectral weights, which specifically includes remote sensing image preprocessing, under-segmentation object production, spectral weight estimation, segmentation scale estimation based on "spectral weights", multi-scale segmentation based on "spectral weights", segmentation result merging, and accuracy verification. Aiming at the problem of low object-oriented scale estimation and classification accuracy, a multi-scale segmentation method for remote sensing images based on spectral weights is proposed, a scale estimation method based on spectral weights to improve the average local variance, and a multi-scale segmentation method based on spectral weights to improve the segmentation scale effectively, which improves the segmentation accuracy of remote sensing images to a certain extent, and is of great significance for remote sensing image classification and information extraction.

[0004] Based on the above scheme, it is found that the current remote sensing image segmentation technology rarely combines task load analysis and parallel allocation rules to optimize the segmentation process in a targeted manner. It lacks refined task allocation and processing strategies, which easily leads to problems such as uneven task allocation, waste of computing resources and excessive computing time, thereby limiting the improvement of segmentation efficiency. In addition, the existing technology fails to fully consider the heterogeneity of different image types and morphological characteristics of objects, which easily leads to inaccurate segmentation results or unreasonable merging when processing complex areas, thereby affecting the segmentation accuracy. The existing regional merging strategy is also relatively simple, ignoring factors such as shape and spatial relationship, which in turn affects the merging effect and the accuracy of the final result. Due to the lack of multi-scale optimized segmentation and efficient scheduling of parallel computing, the existing technology often has a large compromise between segmentation accuracy and computing efficiency, and it is difficult to efficiently solve the segmentation problem of large-scale remote sensing images. Therefore, in practical applications, the existing technology is often difficult to meet the requirements of efficient and accurate remote sensing image segmentation, and it is difficult to achieve comprehensive optimization of performance and efficient identification of potential problems. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a multi-scale object-oriented parallel segmentation method for remote sensing images, which solves the problems of uneven task allocation, waste of computing resources, insufficient segmentation accuracy and inaccurate region merging existing in the prior art.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-scale object-oriented parallel segmentation method for remote sensing images, comprising the following steps: obtaining remote sensing image data to be segmented, and after type judgment processing, performing slicing processing to obtain a number of slice blocks; performing complexity analysis on each slice block to obtain the task load index of each slice block, and performing sorting processing to obtain a slice task allocation sequence; obtaining the number of CPU cores, and according to a preset allocation rule, the slice task allocation sequence is allocated, and image segmentation parallel processing is performed to obtain a number of segmented areas of each slice block; performing region merging processing on the several segmented areas of each slice block respectively, and after the region merging processing, splicing and position coordinate mapping processing are performed on each slice block.

[0007] Furthermore, after the type judgment processing, the specific steps of slicing processing are as follows: if the remote sensing image data to be segmented is a multispectral image, the multispectral image pixels are superimposed to generate a full-color grayscale image; if the remote sensing image data to be segmented is a single-band image or a full-color grayscale image, the GDAL reading function is used to slice the large remote sensing image according to black blocks of a preset size to generate m*n gray slice images, and for slice images whose edges do not meet the preset size, black blocks are used to supplement them.

[0008] Furthermore, the parallel processing of image segmentation is specifically executed in parallel for each process, including: executing gradient calculation processing, edge detection processing (Sobel operator), Gaussian filtering noise reduction processing and marked watershed segmentation processing in parallel for each process.

[0009] Furthermore, the specific steps of performing the marked watershed segmentation process are as follows: read the average gradient amplitude of the slice block and the gradient amplitude of each pixel from the edge detection process result, and perform a comprehensive analysis to obtain the adaptive gradient amplitude of each pixel of the slice block; based on the adaptive gradient amplitude of each pixel of the slice block, randomly select a number of initial seed points; for each initial seed point, analyze the growth index of each adjacent pixel within a preset neighborhood range, and merge the adjacent pixels with the largest growth index with the initial seed point, and update the seed center of the region; repeat the growth index analysis, regional merging, and seed updating steps until the segmentation is completed.

[0010] Furthermore, the specific formula for calculating the growth index of each adjacent pixel within the preset neighborhood range of each initial seed point is as follows: in, is the growth index of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, is the adaptive gradient amplitude of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, is the Euclidean distance value of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, u = 1, 2, 3, ..., u 0 ,u 0 is the number of initial seed points, p = 1, 2, 3, ..., p 0 , p 0 is the number of adjacent pixels within the preset neighborhood.

[0011] Furthermore, the complexity analysis of each slice block is performed to obtain the specific steps of the task allocation complexity correction index of each slice block as follows: read the number of pixels of each slice block, and analyze the preprocessing time of each slice block according to the preset pixel number interval; analyze the grayscale pixel standard deviation of each slice block, and read the average gradient amplitude of each slice block from the edge detection process, and perform a comprehensive analysis in combination with the number of pixels and preprocessing time of each slice block to obtain the task load index of each slice block. The specific formula is as follows: Among them, RfZ i is the task load index of the i-th slice block, XsY i is the number of pixels in the i-th slice block, XsT i PtD is the pixel number adjustment factor of the i-th slice block stored in the database. i is the average gradient amplitude of the i-th slice block, PtT i is the average gradient adjustment factor of the i-th slice block stored in the database, α is the gradient adjustment coefficient stored in the database, HbC i is the grayscale pixel standard deviation of the i-th slice block, HbT i is the grayscale pixel standard deviation adjustment factor of the i-th slice block stored in the database, ω is the grayscale adjustment coefficient stored in the database, YcL i is the preprocessing time of the i-th slice block, YcT i is the time adjustment factor of the i-th slice block stored in the database, μ is the processing time adjustment coefficient stored in the database, i = 1, 2, 3, ..., i 0 ,i 0 The number of slice blocks.

[0012] Furthermore, the specific steps of performing region merging processing on several segmented regions of each slice block are as follows: reading the number of pixels of each segmented region of each slice block, and taking them as the segmented region area value respectively, and performing comparative analysis to obtain the minimum segmented region; judging whether the segmented region area of ​​the minimum segmented region is lower than the preset segmented region area threshold; and when the segmented region area of ​​the minimum segmented region is lower than the preset segmented region area threshold, analyzing the merging index of the minimum segmented region and each adjacent segmented region within a preset range, and performing region merging processing on the minimum segmented region and the adjacent segmented region with the largest merging index until the segmented region area of ​​the merged segmented region meets the expected standard.

[0013] Furthermore, the specific steps of analyzing the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range are: analyzing the Pearson correlation coefficient between the minimum segmentation area and each adjacent segmentation area within the preset range, the grayscale pixel mean and grayscale pixel standard deviation of the minimum segmentation area and each adjacent segmentation area within the preset range, and performing a comprehensive analysis to obtain the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range.

[0014] Furthermore, the specific formula for calculating the merging index of the minimum segmentation area and each adjacent segmentation area within a preset range is as follows: Among them, HbZ r is the merging index of the minimum segmentation area and the rth adjacent segmentation area within the preset range, PxX r is the Pearson correlation coefficient between the minimum segmentation area and the rth adjacent segmentation area within the preset range, ZhJ is the mean grayscale pixel value of the minimum segmentation area, HxS r is the grayscale pixel mean of the rth adjacent segmented area within the preset range, QhX is the grayscale pixel mean of the slice block, θ is the grayscale pixel adjustment factor stored in the database, ZbC is the grayscale pixel standard deviation of the minimum segmented area, HbC r is the grayscale pixel standard deviation between the minimum segmentation area and the rth adjacent segmentation area within the preset range, η is the grayscale pixel influence coefficient stored in the database, KjJ r is the regional spatial distance value between the minimum segmentation area and the rth adjacent segmentation area within the preset range, δ is the spatial distance influence coefficient stored in the database, r = 1, 2, 3, ..., r 0 , r 0 is the number of adjacent segmented areas within the preset range.

[0015] Furthermore, the specific steps of splicing and position coordinate mapping of each slice block are as follows: splicing each slice block according to the m*n matrix rule to obtain the segmented remote sensing image data; obtaining the coordinate system information of the remote sensing image data to be segmented, and mapping the segmented remote sensing image data.

[0016] The present invention has the following beneficial effects:

[0017] (1) This multi-scale object-oriented parallel segmentation method for remote sensing images can effectively reduce processing time by dividing remote sensing image data into multiple slice blocks and performing parallel task allocation. Each slice block is prioritized according to its complexity, thereby ensuring that computing resources are evenly distributed according to task load, avoiding certain complex tasks from slowing down the overall speed. The segmentation process of each slice block is processed in parallel, such as gradient calculation, edge detection, noise processing, etc., which can significantly speed up the execution speed of large-scale image segmentation, especially when rapid response or processing of large-scale remote sensing data is required. It has obvious advantages.

[0018] (2) This multi-scale object-oriented parallel segmentation method for remote sensing images calculates the adaptive gradient amplitude of each pixel and performs region growth analysis during the marked watershed segmentation process. It can flexibly adjust the segmentation boundary according to local image features, avoiding detail loss or incorrect segmentation caused by over-reliance on global information in traditional methods. When merging regions, it considers multiple factors such as the area of ​​each segmented region, the correlation between adjacent regions, the grayscale mean and standard deviation. Especially when processing small regions, by calculating the merging index and the Pearson correlation coefficient of adjacent regions, the scientific nature of the merging decision is guaranteed, ensuring the effective merging of the smallest region, avoiding over-merging or omissions, and optimizing the segmentation effect.

[0019] (3) This multi-scale object-oriented parallel segmentation method for remote sensing images can flexibly process various types of remote sensing image data and adopt different slicing processing strategies for different data types. For example, in multispectral image processing, full-color grayscale image overlay processing is adopted, while for single-band images, GDAL is used for slicing and edge filling, ensuring that all types of image data can be efficiently segmented through corresponding processing methods. At the same time, task load analysis effectively balances the processing load and computing resource utilization by calculating indicators such as the preprocessing time and number of pixels of each slice, thereby improving the overall efficiency of the system.

[0020] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the multi-scale object-oriented parallel segmentation method for remote sensing images of the present invention.

[0022] Figure 2 The present invention is a flowchart of the specific steps of performing the labeled watershed segmentation process in the multi-scale object-oriented parallel segmentation method for remote sensing images.

[0023] Figure 3 The present invention is a flowchart of the specific steps of obtaining the complexity correction index of the task allocation of each slice block in the multi-scale object-oriented parallel segmentation method of remote sensing images. DETAILED DESCRIPTION

[0024] The overall idea of ​​the problem in the embodiment of this application is as follows:

[0025] First, obtain the remote sensing image data to be segmented and make a type judgment. If it is a multispectral image, superimpose it to generate a full-color grayscale image. If it is a single-band image or a full-color grayscale image, use the GDAL reading function to slice and process the image to generate multiple slice blocks. Then, perform complexity analysis on each slice block, calculate the task load index, and sort and allocate tasks according to these indicators. When performing image segmentation, each process executes multiple tasks in parallel, including gradient calculation, edge detection, noise processing, and marked watershed segmentation. This process greatly improves the processing speed through parallelization and ensures the efficiency of large-scale data processing. The segmented areas are merged, and the minimum area is merged with the adjacent area according to the merging index until the preset area threshold is met. Finally, through position coordinate mapping, each slice block is spliced ​​into complete remote sensing image data to ensure the accuracy and integrity of the data.

[0026] See also Figure 1The embodiment of the present invention provides a technical solution: a multi-scale object-oriented parallel segmentation method for remote sensing images, comprising the following steps: obtaining remote sensing image data to be segmented, and after type determination processing, performing slicing processing to obtain a number of slice blocks, and using them as basic units of parallel tasks; performing complexity analysis on each slice block to obtain a task load index of each slice block, and performing sorting processing to obtain a slice task allocation sequence; obtaining the number of CPU cores, and allocating the slice task allocation sequence according to a preset allocation rule (for example: assuming that there are 8 CPU cores, the initial total load of each core is 0, first assign the task with the largest load to the first core, at which time the total load of the first core is the load value of the task, and then assign the task with the second largest load to the one with the smallest current total load among the remaining cores, and repeat this process until all the cores have completed the task, and the task with the second largest load is assigned ... second largest load. The tasks are all assigned), and image segmentation is processed in parallel to obtain several segmented areas of each slice block, wherein the number of CPU cores is read using the MPI parallel function library, and the segmentation scale set by the user needs to be input in advance. The segmentation scale should be based on the actual resolution of the image unit. If the image resolution ratio is 0.5 meters, and the user inputs a segmentation scale of 10, then the smallest segmentation unit Area after image segmentation corresponds to 10*0.5*10*0.5=25 square meters, and each process is executed in parallel, including: first, gradient calculation and edge detection are performed, then Gaussian filtering is used for noise reduction, and finally, the marked watershed algorithm is used for regional segmentation to achieve multi-scale object-oriented segmentation; the several segmented areas of each slice block are respectively merged, and after the regional merging process, each slice block is spliced ​​and the position coordinate mapping process is performed.

[0027] Specifically, after the type judgment processing, the specific steps of slicing processing are as follows: if the remote sensing image data to be segmented is a multispectral image, the multispectral image pixels are superimposed to generate a full-color grayscale image; if the remote sensing image data to be segmented is a single-band image or a full-color grayscale image, the GDAL reading function is used to slice the large remote sensing image according to black blocks of a preset size (for example, 512*512 pixels) to generate m*n gray slice images, and for slice images whose edges do not meet the preset size (for example, the edge does not meet 512 width or height), black blocks are used to supplement them.

[0028] In this implementation, through type judgment processing, the method can adopt different slicing processing strategies according to different remote sensing image types (such as multispectral images, single-band images or full-color grayscale images). For example, for multispectral images, by superimposing their pixels to generate full-color grayscale images, the computational complexity brought by multispectral data can be reduced while retaining the main feature information of the image; for single-band or full-color images, using GDAL function slicing processing can ensure that the method is applicable to various remote sensing data types, improve the versatility and flexibility of processing, and through the preset size (such as 512*512 pixels) bl Block slicing, the size of each slice block is fixed, which helps to reasonably divide the computing tasks and avoids the low computing efficiency caused by excessive memory consumption or too small data blocks. The slicing method makes the loading and processing of large-scale remote sensing image data in memory more efficient, and also helps parallel processing and task allocation, improving computing performance. By filling in the slices whose edges do not meet the preset size, it can avoid affecting the accuracy of segmentation due to incomplete edge data. The filling process ensures the consistency of the size of each slice, so that subsequent parallel processing and image segmentation tasks can proceed smoothly, avoiding inconsistencies or errors caused by missing edge data. The slicing process combines the specific size and segmentation requirements of the image. Through reasonable image segmentation and edge filling processing, it simplifies the data preprocessing steps, avoids complex image cropping and filling operations, and improves the processing efficiency of the system.

[0029] Specifically, the parallel processing of image segmentation is executed in parallel for each process, including: performing gradient calculation processing, edge detection processing (Sobel operator), Gaussian filtering noise reduction processing and marked watershed segmentation processing in parallel for each process, specifically: using the Sobel operator to perform image gradient processing on the section image to generate an edge detection image edge Image; using a 5*5 smoothing filter operator to remove Gaussian white noise from the edge detection image edge Image to generate an image gaussEdgeImage; the slice image segmentation algorithm uses the marked watershed algorithm to perform watershed segmentation on the image gaussEdgeImage, creating a marked image of the same size as the section image for seed point growth, continuously flooding the region through the local minimum value of the image, continuously raising the water surface until the edge of the gaussEdgeImage is encountered, and after all seed points are flooded, a segmentation boundary image formed between different objects in the image is obtained.

[0030] In this embodiment, by dividing the image segmentation process into multiple independent parallel tasks (such as gradient calculation, edge detection, noise removal and watershed segmentation), multi-core CPU or distributed computing architecture can be used for parallel processing. This parallelization method significantly improves the processing speed and is particularly suitable for the processing of large-scale remote sensing image data. The independent execution of each process reduces the computing bottleneck, so that large-scale image segmentation tasks can be completed efficiently. Using the Sobel operator for gradient calculation can efficiently extract edge information in the image, which is a key step in image segmentation. Subsequently, by using Gaussian filtering (5x5 smoothing filter operator) to remove Gaussian white noise, the influence of noise on the segmentation result can be effectively reduced, and the accuracy of edge detection is improved. This process ensures that when the image is segmented, the object boundary can be more accurately identified, and the interference of noise on subsequent analysis is reduced. The marked watershed algorithm is a segmentation method based on image gradient. By simulating the principle of "flooding on the water surface", seed points are created at the local minimum of the image, and regions are distinguished along the gradient height to finally form the boundary between objects. Through this method, different regions in the image can be automatically and accurately identified to form accurate segmentation boundaries, which is particularly suitable for complex, blurred or overlapping objects. The segmentation between bodies can effectively deal with the detail differences between different objects. The labeled watershed algorithm can effectively deal with complex situations in different types of images through local minima and regional flooding. It is especially suitable for images with overlapping or similar backgrounds. The algorithm does not rely on specific image shapes and colors, but is based on gradient information. Therefore, it has strong robustness and adaptability, and can perform well in a variety of environments and conditions. By combining gradient calculation, edge detection, noise removal and watershed algorithm, this method can accurately detect the boundaries of objects in the image, avoiding the fuzzy segmentation or mis-segmentation that may occur in traditional methods, especially in complex images with higher segmentation accuracy, which makes the final segmentation result more in line with actual needs, especially suitable for remote sensing image analysis with high accuracy requirements.

[0031] Specifically, Figure 2As shown in the figure, the specific steps of performing the marked watershed segmentation process are as follows: read the average gradient amplitude of the slice block and the gradient amplitude of each pixel from the edge detection process result, and perform a comprehensive analysis (i.e., gradient amplitude / average gradient amplitude) to obtain the adaptive gradient amplitude of each pixel of the slice block; based on the adaptive gradient amplitude of each pixel of the slice block, randomly select a number of initial seed points (take the several pixels corresponding to the minimum value); for each initial seed point, analyze the growth index of each adjacent pixel within the preset neighborhood range, and merge the adjacent pixels with the largest growth index with the initial seed point, and update the seed center of the region (i.e., the position center point); repeat the growth index analysis, region merging, and seed updating steps until the segmentation is completed, that is, until an obvious edge is encountered (i.e., the gradient rises suddenly or the growth index drops below a certain threshold, thereby terminating the expansion of the region).

[0032] The specific formula for calculating the growth index of each adjacent pixel within the preset neighborhood of each initial seed point is as follows: in, is the growth index of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, is the adaptive gradient amplitude of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, is the Euclidean distance value of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, u = 1, 2, 3, ..., u 0 ,u 0 is the number of initial seed points, p = 1, 2, 3, ..., p 0 , p 0 is the number of adjacent pixels within the preset neighborhood.

[0033] In this implementation, the adaptive gradient amplitude is calculated based on the ratio of the gradient amplitude of each pixel to the average gradient amplitude of the slice block, and the gradient weight can be dynamically adjusted according to the local image features. This method avoids the limitations brought by the global fixed threshold, so that the gradient information of different regions in the image can be adaptively adjusted according to the local situation, thereby improving the accuracy of edge detection and the detail retention of the segmentation results. By randomly selecting multiple initial seed points and analyzing the growth potential of adjacent pixels based on the growth index, the selection of seed points is made more intelligent and dynamic. The calculation of the growth index takes into account the adaptive gradient amplitude and the Euclidean distance, which can ensure that the seed points grow to the most representative areas and avoid the erroneous merging or missing areas that may be caused by relying solely on the traditional seed point selection method. This method makes the regional merging more consistent with the actual structure of the image and reduces the risk of erroneous merging during the segmentation process. When performing regional growth, as the seed points expand to adjacent pixels, the conditions of a sudden increase in the gradient or a decrease in the growth index are used to determine whether to terminate the regional expansion. This method can flexibly respond to different image features and avoid the inaccurate or over-expanded boundaries caused by fixed thresholds in traditional methods. By dynamically adjusting the segmentation boundaries, the boundaries between objects in the image can be more accurately identified, especially in complex or blurred areas. Since this method integrates gradient information and neighborhood growth index in the segmentation process, it can better adapt to images of different types and complexities, whether it is processing relatively simple images or images with complex textures and boundaries. By integrating multiple image features, the robustness of segmentation can be improved, ensuring that attention is paid to details and over-simplification is avoided in various situations. By expanding and merging regions through the marked watershed algorithm, this method can handle details in large-scale images and avoid situations where it is difficult to effectively distinguish in complex areas. At the same time, through the continuous adjustment of seed point updates and region merging, the algorithm can continuously optimize the segmentation results to ensure that the final segmentation effect conforms to the actual structure of the image content. Especially when processing high-resolution big data such as remote sensing images, it demonstrates its efficiency and accuracy.

[0034] Specifically, Figure 3 As shown in FIG. 1 , the specific steps of performing complexity analysis on each slice block and obtaining the task allocation complexity correction index of each slice block are as follows: reading the number of pixels of each slice block, and analyzing the preprocessing time of each slice block according to the preset pixel number interval, that is, each preset pixel number interval corresponds to a preprocessing time; analyzing the grayscale pixel standard deviation of each slice block, and reading the average gradient amplitude of each slice block from the edge detection process, and performing a comprehensive analysis in combination with the number of pixels and preprocessing time of each slice block to obtain the task load index of each slice block, and the specific formula is as follows: Among them, RfZ i is the task load index of the i-th slice block, XsY iis the number of pixels in the i-th slice block, XsT i is the pixel number adjustment factor of the i-th slice block stored in the database (it is the reference pixel number used for normalization, for example, for a standard 512*512 block, the value can be 512*512), PtD i is the average gradient amplitude of the i-th slice block, PtT i is the average gradient adjustment factor of the i-th slice block stored in the database (which is a normalization constant used to reflect the typical gradient value), α is the gradient adjustment coefficient stored in the database, HbC i is the grayscale pixel standard deviation of the i-th slice block, HbT i is the grayscale pixel standard deviation adjustment factor of the i-th slice block stored in the database (which is a normalization constant), ω is the grayscale adjustment coefficient stored in the database, YcL i is the preprocessing time of the i-th slice block, YcT i is the time adjustment factor of the i-th slice block stored in the database (which is the historical average processing time of several slice blocks with the same number of pixels), μ is the processing time adjustment coefficient stored in the database, i = 1, 2, 3, ..., i 0 ,i 0 The number of slice blocks.

[0035] It should be explained that the specific steps for obtaining the gradient adjustment coefficient α stored in the database are as follows: the gradient adjustment coefficient is usually obtained by analyzing the image gradient information stored in the database. Specifically, first, a data set containing image block information is extracted from the database, and the gradient value of the image block is calculated (edge ​​detection may be performed using methods such as the Sobel operator). The gradient amplitudes of these image blocks are statistically analyzed to obtain their normalization factors (i.e., normalization constants) to ensure that the gradient differences between different image blocks can be processed uniformly. This coefficient is used to adjust the gradient information of the image so that the characteristics of the image can be better reflected in subsequent calculations.

[0036] The specific steps for obtaining the grayscale adjustment coefficient ω stored in the database are as follows: the acquisition of this coefficient is also based on the grayscale characteristics of the image block. First, the stored image block data is read from the database, and the grayscale pixel value and standard deviation of each image block are calculated. Based on these statistical information, a standardized grayscale adjustment coefficient is generated. This coefficient can adjust the difference in grayscale values ​​between image blocks, ensuring that the image blocks have a consistent grayscale standard during processing, and avoiding the impact of excessive grayscale differences on subsequent image analysis.

[0037] The specific steps for obtaining the processing time adjustment coefficient μ stored in the database are as follows: the processing time adjustment coefficient is based on the historical data of image block processing. First, the processing time data of historical image blocks are saved in the database. These data reflect the time required to process these blocks under different conditions. By analyzing these historical processing time data, an average value is calculated, and the processing time adjustment coefficient is obtained by normalizing the data. This coefficient is used to adjust the processing time of each image block, so that when performing large-scale image processing, the processing time of each block can be estimated more accurately to optimize the processing efficiency.

[0038] In this implementation, the task load index of each slice block is calculated through complexity analysis, and computing resources can be reasonably allocated to each slice block. Specifically, the task load index combines multiple factors such as the number of pixels, preprocessing time, grayscale pixel standard deviation and gradient amplitude of the slice block, and can accurately evaluate the computational complexity of each slice block, so as to dynamically adjust resource allocation according to the actual load situation. This method can avoid excessive waste or shortage of resources and ensure maximum efficiency in the processing process. By analyzing the processing time of each slice block according to the preset relationship between the number of pixels and the preprocessing time, the processing time of complex slice blocks can be estimated in advance. When multiple slices are processed simultaneously, the computing resources can be intelligently scheduled according to the task load index to make more efficient use of the available CPU or GPU cores and avoid excessive backlog of certain tasks, which may lead to delays. The calculation method of the task load index comprehensively considers multiple indicators such as the number of pixels, gradient amplitude, and grayscale standard deviation, making the complexity evaluation of each slice more accurate and comprehensive. This comprehensive analysis not only helps to predict the processing time more accurately, but also can dynamically adjust the actual complexity of the slice, making the processing process more flexible and efficient. For example, in more complex areas (such as Areas with large gradient amplitudes or obvious grayscale changes will get more computing resources, while simple areas can be processed more quickly. By introducing factors such as adjustment factors, normalization constants, and historical processing time, it is possible to better avoid unbalanced computing resource allocation caused by differences in size, grayscale, gradient, etc. between different slices. For example, adjusting the processing time through historical data can accurately reflect the actual computing needs of similar slices in the past, so that similar slices in the future can get appropriate processing time allocation, avoiding excessive allocation of computing resources to some areas that do not require complex processing. When processing large-scale remote sensing images, the number and complexity of slices are often very large. Through the comprehensive evaluation of the task load index, reasonable resource scheduling and task priority sorting can be made in the huge data processing to ensure that each slice gets the appropriate processing time and resources, thereby improving the overall processing capacity of the system and avoiding bottlenecks or resource conflicts in large-scale data processing. Through precise load analysis and resource scheduling, the system can effectively cope with image data of different complexities, ensuring that the system will not crash or the processing time will be too long due to excessive load of a single slice when performing image segmentation tasks. The scalability of this method also makes it adaptable to increasingly complex remote sensing image data sets.

[0039] Specifically, the specific steps of performing region merging processing on several segmented regions of each slice block are as follows: reading the number of pixels of each segmented region of each slice block, and taking them as the segmented region area value respectively, and performing comparative analysis to obtain the minimum segmented region (the segmented region area value is the smallest); judging whether the segmented region area of ​​the minimum segmented region is lower than the preset segmented region area threshold; and when the segmented region area of ​​the minimum segmented region is lower than the preset segmented region area threshold, analyzing the merging index of the minimum segmented region and each adjacent segmented region within a preset range, and performing region merging processing on the minimum segmented region and the adjacent segmented region with the largest merging index until the segmented region area of ​​the merged segmented region meets the expected standard (that is, the segmented region area is equal to or higher than the preset segmented region area threshold).

[0040] The Pearson correlation coefficient between the minimum segmentation area and each adjacent segmentation area within the preset range, the grayscale pixel mean and grayscale pixel standard deviation of the minimum segmentation area and each adjacent segmentation area within the preset range are analyzed, and a comprehensive analysis is performed to obtain the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range.

[0041] The specific formula for calculating the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range is as follows: Among them, HbZ r is the merging index of the minimum segmentation area and the rth adjacent segmentation area within the preset range, PxX r is the Pearson correlation coefficient between the minimum segmentation area and the rth adjacent segmentation area within the preset range, ZhJ is the mean grayscale pixel value of the minimum segmentation area, HxS r is the grayscale pixel mean of the rth adjacent segmented area within the preset range, QhX is the grayscale pixel mean of the slice block, θ is the grayscale pixel adjustment factor stored in the database, ZbC is the grayscale pixel standard deviation of the minimum segmented area, HbC r is the grayscale pixel standard deviation between the minimum segmentation area and the rth adjacent segmentation area within the preset range, η is the grayscale pixel influence coefficient stored in the database, KjJ r is the regional spatial distance value between the minimum segmentation area and the rth adjacent segmentation area within the preset range (i.e., the distance value between the center points of two regions), δ is the spatial distance influence coefficient stored in the database, r = 1, 2, 3, ..., r 0 , r 0 is the number of adjacent segmented areas within the preset range.

[0042] It should be explained that the specific steps for obtaining the grayscale pixel adjustment factor θ stored in the database are as follows: the grayscale image adjustment coefficient is obtained through statistical analysis of the image data stored in the database. First, the grayscale information of the image is used to statistically calculate the grayscale distribution between various areas of the image. The coefficient is used to standardize the grayscale features of the image. By analyzing the correlation of grayscale changes in historical image data, the standardization factor is determined to adjust the difference between image grayscales, so that the results of image segmentation or other subsequent processing are more accurate. The coefficient is usually obtained through the grayscale information of the image block and adjusted according to the distribution of the data.

[0043] The specific steps for obtaining the grayscale pixel influence coefficient η stored in the database are as follows: the grayscale pixel influence coefficient is calculated based on the grayscale differences in different areas of the image. The influence coefficient can be obtained by statistically analyzing the grayscale pixel standard deviation and grayscale influence characteristics of historical image blocks. First, a representative image block is obtained from the database, the standard deviation of its grayscale pixels is calculated, and the data is normalized. Then, by analyzing the grayscale differences of multiple image blocks, the degree of influence of grayscale pixels on regional processing is obtained, and finally the coefficient is determined.

[0044] The specific steps for obtaining the spatial distance influence coefficient δ stored in the database are as follows: the spatial distance influence coefficient is used to adjust the spatial distance influence between image regions. In spatial analysis, regions with a larger distance may have less influence on each other, so adjustment is required. The coefficient is obtained by analyzing the spatial distance between image regions, extracting the spatial information of image blocks from the database, calculating the spatial distance between each pair of regions, and performing normalization processing based on these distance data to finally obtain the spatial distance influence coefficient. This coefficient reflects the degree of influence of spatial position on image regions, which is particularly important in the segmentation process and is used to adjust the influence of regions with a larger distance on the processing process.

[0045] In this embodiment, by comparing and merging the segmented areas of each slice block, unreasonable detail segmentation or noise caused by too small segmented areas can be avoided. Specifically, the merging process ensures that the merged area reaches the expected size by setting an area threshold, avoiding the residue of too small areas, which helps to improve the overall quality of segmentation, making the final segmented area more reasonable and consistent with the actual object structure. When merging the minimum segmented area, multiple factors are considered, such as the Pearson correlation coefficient, the grayscale pixel mean, the grayscale pixel standard deviation, etc. The comprehensive analysis of these factors can ensure the rationality of the merging. The grayscale pixel mean and standard deviation can effectively judge the similarity of adjacent areas, the Pearson correlation coefficient measures the correlation between two areas, and the regional spatial distance takes into account the influence of physical position. Through these multi-dimensional analyses, the algorithm can more intelligently determine which areas should be merged, thereby avoiding mismerging or erroneous segmentation, ensuring the accuracy and reliability of the segmentation results, and quantifying the similarity between the minimum segmented area and the adjacent area through the merging index. This method can dynamically judge the possibility of merging adjacent areas, thereby ensuring that the segmentation boundary conforms to the image content. Actual structure, for example, grayscale pixel mean and standard deviation help identify similar areas, while spatial distance considerations help determine areas that may physically belong to the same object. In this way, the influence of mis-segmentation and noise can be effectively reduced, and the image segmentation quality can be improved. By setting the merging criteria (such as the area of ​​the merged region reaches or exceeds the threshold), the granularity of segmentation can be flexibly controlled in actual operation to avoid excessive merging or too many small regions remaining. This mechanism can be adjusted as needed to adapt to the segmentation requirements in different scenarios, and can achieve more refined and efficient segmentation when processing complex images. Region merging not only considers a single region, but also considers the relationship with adjacent regions. By combining multiple factors, it ensures that each merged region can meet the image segmentation requirements in terms of local consistency and overall structure, which makes the segmented region closer to the real object boundary in the image, thereby improving the overall quality of the image. By automatically calculating the merging index of each slice block and performing region merging according to these indices, the complexity of manual intervention is avoided, which significantly improves the efficiency of large-scale image processing, and can automatically adjust the merging criteria to make the processing more efficient and accurate.

[0046] Specifically, the specific steps of splicing and position coordinate mapping of each slice block are as follows: splicing each slice block according to the m*n matrix rule, merging each splicing position after determination using the Pearson coefficient, removing the segmentation lines of homogeneous regions of adjacent segmented images, and obtaining segmented remote sensing image data; obtaining the coordinate system information of the remote sensing image data to be segmented, and mapping the segmented remote sensing image data to form the final result, including: mapping the entire marked data with the remote sensing image, reading the remote sensing image coordinate system, and performing position coordinate mapping on the entire marked data so that the marked data is aligned with the remote sensing image.

[0047] In this implementation, by splicing the slice blocks according to the m*n matrix rule and using the Pearson correlation coefficient to determine the similarity of adjacent regions, it is possible to accurately determine whether adjacent slice blocks belong to the same object or region. This method can not only achieve correct splicing of regions, but also effectively remove the homogeneous region segmentation lines between adjacent segmented images, thereby avoiding erroneous overlap or inconsistency in the splicing process. This makes the spliced ​​remote sensing image data more natural, reduces the error on the segmentation boundary, and improves the overall quality of image segmentation. By removing the homogeneous region segmentation lines of adjacent segmented images, the image segmentation results can be further refined and optimized. Especially when the features of adjacent regions are similar, traditional methods may have obvious splicing marks or false boundaries. By using the Pearson coefficient as a reference, the splicing method can be dynamically adjusted to make the segmented image smoother and more in line with the boundaries of real objects. This optimization can effectively improve the accuracy of image processing and make the image segmentation results more in line with actual needs. In remote sensing image data processing, coordinates System mapping is a crucial step. By obtaining the coordinate system information of the remote sensing image to be segmented and performing position coordinate mapping, it can ensure that the final segmented data is perfectly aligned with the original remote sensing image. This is particularly important when performing spatial analysis and subsequent processing. It can avoid coordinate misalignment or data errors, thereby ensuring a high degree of accuracy when performing further analysis (such as measurement, identification, etc.). Through the steps of automatic stitching and coordinate system mapping, this method can efficiently stitch multiple slice blocks into complete remote sensing image data while ensuring the accuracy of data coordinates, which greatly improves processing efficiency. Especially when processing large-scale remote sensing data, there is no need to manually adjust slices or coordinates, which greatly saves time and labor costs. The final stitching results and coordinate mapping ensure that the segmented remote sensing image data is not only visually consistent, but also spatially correct, which makes the subsequent analysis based on the segmentation results more accurate and reliable. For example, when applied in fields such as environmental monitoring and land use change analysis, more accurate geographic spatial information can be obtained.

[0048] In summary, this application has at least the following effects:

[0049] By dividing the remote sensing image data into multiple slice blocks and performing parallel task allocation, the processing time can be effectively reduced. Each slice block is prioritized according to its complexity, thereby ensuring that computing resources are evenly distributed according to task load, avoiding certain complex tasks from slowing down the overall speed. Parallel processing of the segmentation process of each slice block, such as gradient calculation, edge detection, noise processing, etc., can significantly speed up the execution speed of large-scale image segmentation, especially when rapid response or processing of large-scale remote sensing data is required, it has obvious advantages.

[0050] By calculating the adaptive gradient amplitude of each pixel and performing regional growth analysis during the marked watershed segmentation process, the segmentation boundary can be flexibly adjusted according to local image features, avoiding detail loss or incorrect segmentation caused by traditional methods that rely too much on global information. When merging regions, multiple factors such as the area of ​​each segmented region, the correlation of adjacent regions, the grayscale mean and standard deviation are considered. Especially when processing small regions, the scientific nature of the merging decision is guaranteed by calculating the merging index and the Pearson correlation coefficient of adjacent regions, ensuring the effective merging of the smallest area, avoiding excessive merging or omissions, and optimizing the segmentation effect.

[0051] It can flexibly process various types of remote sensing image data and adopt different slicing processing strategies for different data types. For example, in multispectral image processing, full-color grayscale image overlay processing is adopted, while for single-band images, GDAL is used for slicing and edge filling, ensuring that all types of image data can be efficiently segmented through corresponding processing methods. At the same time, task load analysis effectively balances the processing load and computing resource utilization by calculating indicators such as the preprocessing time and number of pixels of each slice, thereby improving the overall efficiency of the system.

[0052] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0053] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-scale object-oriented parallel segmentation method for remote sensing images, characterized in that: The following steps are involved: Acquire the remote sensing image data to be segmented, and after type determination, perform slicing to obtain a number of slice blocks; Perform complexity analysis on each slice block to obtain the task load index of each slice block, and perform sorting to obtain the slice task allocation sequence; Obtain the number of CPU cores, and allocate the slice task allocation sequence according to the preset allocation rules, and perform image segmentation parallel processing to obtain several segmentation areas of each slice block; The plurality of segmented regions of each slice block are respectively subjected to region merging processing, and after the region merging processing, the slice blocks are spliced ​​and subjected to position coordinate mapping processing.

2. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 1, characterized in that: After the type judgment process, the specific steps for slicing are as follows: If the remote sensing image data to be segmented is a multispectral image, the multispectral image pixels are superimposed to generate a full-color grayscale image; If the remote sensing image data to be segmented is a single-band image or a full-color grayscale image, use the GDAL reading function to slice the large remote sensing image according to blocks of a preset size to generate m*n gray slice images, and use blocks to fill in the slice images whose edges do not meet the preset size.

3. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 1, characterized in that: The parallel processing of image segmentation is specifically executed in parallel for each process, including: executing gradient calculation processing, edge detection processing, Gaussian filtering noise reduction processing and marked watershed segmentation processing in parallel for each process.

4. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 3, characterized in that: The specific steps to perform the labeled watershed segmentation process are as follows: The average gradient amplitude of the slice block and the gradient amplitude of each pixel are read from the edge detection processing result, and a comprehensive analysis is performed to obtain the adaptive gradient amplitude of each pixel of the slice block; Based on the adaptive gradient amplitude of each pixel of the slice block, several initial seed points are randomly selected; For each initial seed point, the growth index of each adjacent pixel in the preset neighborhood is analyzed respectively, and the adjacent pixel with the largest growth index is regionally merged with the initial seed point, and the seed center of the region is updated; Repeat the growth index analysis, region merging, and seed updating steps until the segmentation is completed.

5. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 4, characterized in that: The specific formula for calculating the growth index of each adjacent pixel within the preset neighborhood of each initial seed point is as follows: in, is the growth index of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, is the adaptive gradient amplitude of the pth adjacent pixel within the preset neighborhood of the uth initial seed point, is the Euclidean distance value of the pth adjacent pixel in the preset neighborhood range of the uth initial seed point, u=1, 2, 3, …, u0, u0 is the number of initial seed points, p=1, 2, 3, …, p0, p0 is the number of adjacent pixels in the preset neighborhood range.

6. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 1, characterized in that: The specific steps of performing complexity analysis on each slice block and obtaining the task allocation complexity correction index of each slice block are as follows: Read the number of pixels of each slice block, and analyze the preprocessing time of each slice block according to the preset pixel number interval; The standard deviation of the grayscale pixels of each slice block is analyzed, and the average gradient amplitude of each slice block is read from the edge detection process. Combined with the number of pixels and preprocessing time of each slice block, a comprehensive analysis is performed to obtain the task load index of each slice block. The specific formula is as follows: Among them, RfZ i is the task load index of the i-th slice block, XsY i is the number of pixels in the i-th slice block, XsT i PtD is the pixel number adjustment factor of the i-th slice block stored in the database. i is the average gradient amplitude of the i-th slice block, PtT i is the average gradient adjustment factor of the i-th slice block stored in the database, α is the gradient adjustment coefficient stored in the database, HbC i is the grayscale pixel standard deviation of the i-th slice block, HbT i is the grayscale pixel standard deviation adjustment factor of the i-th slice block stored in the database, ω is the grayscale adjustment coefficient stored in the database, YcL i is the preprocessing time of the i-th slice block, YcT i is the time adjustment factor of the i-th slice block stored in the database, μ is the processing time adjustment coefficient stored in the database, i=1, 2, 3, …, i0, i0 is the number of slice blocks.

7. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 1, characterized in that: The specific steps of merging the several segmented regions of each slice block are as follows: Read the number of pixels in each segmented area of ​​each slice block and use them as the area value of the segmented area respectively, and perform comparative analysis to obtain the minimum segmented area; Determine whether the segmentation area of ​​the minimum segmentation area is lower than a preset segmentation area threshold; When the segmentation area of ​​the minimum segmentation area is lower than the preset segmentation area threshold, the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range is analyzed, and the minimum segmentation area and the adjacent segmentation area with the largest merging index are merged until the segmentation area area of ​​the merged segmentation area meets the expected standard.

8. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 7, characterized in that: The specific steps of analyzing the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range are as follows: analyzing the Pearson correlation coefficient between the minimum segmentation area and each adjacent segmentation area within the preset range, the grayscale pixel mean and grayscale pixel standard deviation of the minimum segmentation area and each adjacent segmentation area within the preset range, and performing a comprehensive analysis to obtain the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range.

9. The multi-scale object-oriented parallel segmentation method for remote sensing images according to claim 8, characterized in that: The specific formula for calculating the merging index of the minimum segmentation area and each adjacent segmentation area within the preset range is as follows: Among them, HbZ r is the merging index of the minimum segmentation area and the rth adjacent segmentation area within the preset range, PxX r is the Pearson correlation coefficient between the minimum segmentation area and the rth adjacent segmentation area within the preset range, ZhJ is the mean grayscale pixel value of the minimum segmentation area, HxS r is the grayscale pixel mean of the rth adjacent segmented area within the preset range, QhX is the regional grayscale pixel mean of the slice block, θ is the grayscale pixel adjustment factor stored in the database, ZbC is the grayscale pixel standard deviation of the minimum segmented area, HbC r is the grayscale pixel standard deviation between the minimum segmentation area and the rth adjacent segmentation area within the preset range, η is the grayscale pixel influence coefficient stored in the database, KjJ r is the regional spatial distance value between the minimum segmentation area and the rth adjacent segmentation area within the preset range, δ is the spatial distance influence coefficient stored in the database, r=1, 2, 3, ..., r0, r0 is the number of adjacent segmentation areas within the preset range.

10. The remote sensing image multi-scale object-oriented parallel segmentation method according to claim 1, characterized in that: The specific steps for splicing each slice block and mapping the position coordinates are as follows: Each slice block is spliced ​​according to the m*n matrix rule to obtain the segmented remote sensing image data; The coordinate system information of the remote sensing image data to be segmented is obtained, and the segmented remote sensing image data is mapped.

Images