Large target processing method based on remote sensing image partitioning

Abstract

The invention provides a large-scale target processing method based on remote sensing image blocking. The method comprises the following steps: reading an image and blocking the image; traversing the blocks for target detection; screening the detection frames close to the boundary; calculating the global center coordinates of the blocks, and re-intercepting the blocks for supplementary detection; and outputting a combined result. The method solves the problem of information loss caused by boundary segmentation of the large target during block detection of the high-resolution remote sensing image, remarkably improves the detection integrity and accuracy, and is suitable for detection of the large target such as a ship in a satellite image. The invention relates to the technical field of remote sensing.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a large-scale target processing method based on remote sensing image segmentation. Background Art

[0002] Remote sensing image target detection technology has been widely used in fields such as resource surveys and environmental monitoring. With the development of high-resolution remote sensing satellites, the amount of image data has grown exponentially. Traditional target detection models are limited by computing resources and cannot directly process entire scene images. Therefore, the industry generally adopts image segmentation processing methods, which divide large-scale images into multiple sub-regions for separate detection. To ensure target integrity, existing technologies usually set overlapping areas when segmenting. However, for large targets (such as ships and aircraft), they occupy a large area in the image and are easily segmented multiple times at the block boundaries, resulting in detection results that only cover local areas of the target.

[0003] Although the problem can be alleviated to a certain extent by adjusting the block size and overlap threshold, it cannot fundamentally solve the defect of target information loss at the boundary. Current improvement methods (such as multi-scale feature fusion technology) can improve target detection accuracy, but they are mainly suitable for small-scale image processing and rely on complex model structure adjustments. It is difficult to meet the complete detection requirements of multi-scale targets in large-scale high-resolution remote sensing images. Existing methods are difficult to balance processing efficiency and large-scale target detection integrity, resulting in a decrease in the accuracy of detection results. A large amount of manual correction is required in the later stage, which significantly increases the workload and task cycle.

[0004] Therefore, the inventors urgently need a large-scale target processing method based on remote sensing image segmentation to solve the information loss problem caused by boundary target segmentation and improve the accuracy and usability of detection results. Summary of the Invention

[0005] In response to the above-mentioned defects of the existing technology, the present invention provides a large-scale target processing method based on remote sensing image segmentation, which aims to solve the problems in the existing technology of large-scale target information loss and incomplete detection results due to boundary segmentation during high-resolution remote sensing image segmentation detection, and the difficulty of the existing method in balancing processing efficiency and detection accuracy.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a large-scale target processing method based on remote sensing image segmentation, characterized in that it includes the following steps:

[0007] S1: Read high-resolution remote sensing images, obtain information about the number of rows and columns, set the block size block_size and the overlap threshold overlop_threshold, calculate the block step length stride, and block the remote sensing image according to the block size block_size, the block step length stride, the number of rows and the number of columns cols;

[0008] S2: Traverse each block area generated in step S1, call the optimal weight of the trained target detection model to perform target detection on the current block area, and generate a detection result including the target detection box and its category, location and confidence level;

[0009] S3: traverse each detection frame generated in the current block area in step S2, and determine whether the detection frame is close to the boundary of the current block area;

[0010] S4: For the detection frame determined in step S3 to be close to the boundary of the current block area, calculate the center pixel coordinates (center_x, center_y) of the detection frame, and convert the center pixel coordinates to the global center pixel coordinates (global_center_x, global_center_y) relative to the entire remote sensing image;

[0011] S5: With the global center pixel coordinates (global_center_x, global_center_y) calculated in step S4 as the center, a rectangular area with the same size as the initial block is re-cut on the entire remote sensing image. If the new block area exceeds the image boundary, only the valid area is retained; the optimal weight of the trained target detection model is called to perform target detection on the new block area again to generate a supplementary detection result;

[0012] S6: Combine the detection result of step S2 and the supplementary detection result of step S5, and output the final target detection result to a storage medium.

[0013] Based on the above, a large-scale target processing method based on remote sensing image segmentation has the beneficial effect of solving the problems of large-scale target information loss and incomplete detection results caused by boundary segmentation in the existing high-resolution remote sensing image segmentation detection technology, as well as the difficulty of balancing processing efficiency and detection accuracy in the existing methods. The main advantages are:

[0014] 1. In step S3, the present invention traverses each detection frame and determines whether it is close to the boundary of the current block area. Based on the preset boundary screening logic, the target detection frame located at the edge of the block and at risk of information loss is identified, providing a clear target object for subsequent targeted processing;

[0015] 2. Through the operations of calculating the central pixel coordinates for the detection boxes near the boundary in step S4 and converting them into global central pixel coordinates, and the operation of re - intercepting a rectangular area with the same size as the initial block size centered on the global central coordinate and calling the target detection model for detection again in step S5, after identifying potential incomplete targets, a complete image area that is not cut by the boundary is reconstructed centered on the target itself for detection. This enables the target detection model to have the opportunity to completely detect the target, thus effectively solving the problems of large - target information loss and incomplete detection results caused by block - boundary segmentation.

[0016] 3. By implementing the supplementary detection strategy of steps S4 and S5 only on the detection boxes near the boundary screened in step S3, it is ensured that the vast majority of detection boxes far from the boundary with reliable detection results only need to be detected once, and only a small number of high - risk targets near the boundary are detected twice. While significantly improving the integrity of large - target detection, it effectively controls the overall computational cost and avoids the huge computational burden brought by global repeated detection.

[0017] 4. Through the initial detection of each initial block in step S2, combined with the target - centered supplementary detection of the screened boundary targets in step S5, and by merging the initial detection results and the supplementary detection results in step S6, the capabilities of existing high - performance target detection models are fully utilized. Based on the initial detection, this process accurately supplements the detection of targets in the key link of the block - boundary area and fuses the results, significantly improving the detection accuracy and result reliability for various targets, especially large targets that are vulnerable to block division, while maintaining the high - efficiency block - processing framework.

[0018] Further, in step S1, the formula for calculating the block step size stride is:

[0019] stride = block_size×(1 - overlop_threshold)

[0020] Further, step S3 further includes:

[0021] Obtain the actual width nXBK pixels and actual height nYBK pixels of the current block area. Let j be the horizontal starting coordinate of the current block, i be the vertical starting coordinate of the current block, and k be the boundary screening coefficient. k is a preset constant and satisfies 0 < k < 1. Then:

[0022] When the block is in the non - boundary area of the image, nXBK = block_size, nYBK = block_size;

[0023] When the block is on the right boundary of the image, nXBK = cols - j;

[0024] When the block is located at the lower boundary of the image, nYBK = rows-i;

[0025] If j+block_size≤cols, then nXBK=block_size; otherwise nXBK=cols-j;

[0026] If i + block_size ≤ rows, then nYBK = block_size; otherwise nYBK = rows - i;

[0027] Calculate the screening boundary value:

[0028] Calculate the upper boundary threshold threshold_boundary_y_min=k×block_size;

[0029] Calculate the lower boundary threshold threshold_boundary_y_max=(1-k)×nYBK;

[0030] Calculate the left boundary threshold threshold_boundary_x_min=k×block_size;

[0031] Calculate the right boundary threshold threshold_boundary_x_max=(1-k)×nXBK;

[0032] Let m represent the vertex number of the detection box, m = 1, 2, 3, 4. For the four vertex coordinates (x_m, y_m) of the current detection box, if any vertex coordinate does not satisfy the following relationship:

[0033] threshold_boundary_x_min <x_m<threshold_boundary_x_max

[0034] and

[0035] threshold_boundary_y_min <y_m<threshold_boundary_y_max

[0036] It is determined that the detection frame is close to the boundary of the current block area.

[0037] Furthermore, in step S4, the formula for converting the center pixel coordinates (center_x, center_y) of the detection frame into the global center pixel coordinates (global_center_x, global_center_y) relative to the entire remote sensing image is:

[0038] global_center_x=center_x+j;

[0039] global_center_y=center_y+i.

[0040] Furthermore, in step S5, the boundary coordinates of the newly segmented area are determined in the following manner:

[0041] Calculate the left margin:

[0042] new_block_x_min=max{global_center_x-block_size÷2,0};

[0043] Calculate the upper bound:

[0044] new_block_y_min=max{global_center_y-block_size÷2,0};

[0045] Calculate the right boundary:

[0046] new_block_x_max=min{global_center_x+block_size÷2,cols};

[0047] Calculate the lower bound:

[0048] new_block_y_max=min{global_center_y+block_size÷2,rows}.

[0049] Furthermore, in step S6, the outputting of the final target detection result includes associating and storing the category information, location coordinates and confidence level of each target.

[0050] Furthermore, the position coordinates include the coordinate values ​​of four vertices of the target frame.

[0051] Furthermore, the target detection model is a detection model that outputs a rotated target frame.

[0052] Furthermore, the high-resolution remote sensing image is a satellite remote sensing image, and the large target includes a ship.

[0053] In order to more clearly illustrate the above features of the present invention and the objects to be achieved, the present invention will be further described below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 :It is the flow chart of the present invention;

[0055] Figure 2 : is a schematic diagram of the image segmentation effect of the present invention;

[0056] Figure 3 : This is a schematic diagram of the target detection effect of the block implementation of the present invention;

[0057] Figure 4 : A schematic diagram of the changing effect of the supplementary detection process of the present invention;

[0058] Figure 5 : A true color map of the study area provided by an embodiment of the present invention;

[0059] Figure 6 : This is a rendering of the study area provided in an embodiment of the present invention before being processed by the method;

[0060] Figure 7 : This is a diagram showing the effect of the study area provided in an embodiment of the present invention after being processed by the method of the present invention. DETAILED DESCRIPTION

[0061] See also Figure 1-Figure 7 As shown,

[0062] This embodiment discloses a large-scale target processing method based on remote sensing image segmentation, which is characterized by comprising the following steps:

[0063] S1: Read high-resolution remote sensing images, obtain information about the number of rows and columns, set the block size block_size and the overlap threshold overlop_threshold, calculate the block step length stride, and block the remote sensing image according to the block size block_size, the block step length stride, the number of rows and the number of columns cols;

[0064] S2: Traverse each block area generated in step S1, call the optimal weight of the trained target detection model to perform target detection on the current block area, and generate a detection result including the target detection box and its category, location and confidence level;

[0065] S3: traverse each detection frame generated in the current block area in step S2, and determine whether the detection frame is close to the boundary of the current block area;

[0066] S4: For the detection frame determined in step S3 to be close to the boundary of the current block area, calculate the center pixel coordinates (center_x, center_y) of the detection frame, and convert the center pixel coordinates to the global center pixel coordinates (global_center_x, global_center_y) relative to the entire remote sensing image;

[0067] S5: Taking the global center pixel coordinates (global_center_x, global_center_y) calculated in step S4 as the center, intercept a rectangular area with the same size as the initial block on the entire remote sensing image. If the new block area exceeds the image boundary, only the valid area is retained; call the best weights of the trained object detection model to perform object detection on the new block area again, and generate supplementary detection results;

[0068] S6: Merge the detection results of step S2 and the supplementary detection results of step S5, and output the final object detection results to the storage medium.

[0069] In this embodiment, step S1 specifically includes: reading the input GF2 three-band eight-bit high-resolution image, obtaining the number of rows (rows), the number of columns (cols), the geotransformation parameters, and the projection information. Set the block size (block_size) to 1024×1024, the overlap threshold (overlop_threshold) to 0.3, calculate the block stride (stride), and perform block processing by traversing the image row and column coordinates. The starting position i represents the vertical coordinate, and j represents the horizontal coordinate. For the area at the image edge that is less than the block size, the remaining valid area is retained, and the actual width nXBK and height nYBK of each block are calculated considering the boundary.

[0070] In step S1 of this embodiment, the formula for calculating the block stride stride is:

[0071] stride = block_size × (1 - overlop_threshold).

[0072] In this embodiment, step S3 further includes:

[0073] Obtain the actual width nXBK pixels and the actual height nYBK pixels of the current block area. Let j be the starting horizontal coordinate of the current block, i be the starting vertical coordinate of the current block, and k be the boundary screening coefficient. k is a preset constant and satisfies 0 < k < 1. Then:

[0074] When the block is in the non-boundary area of the image, nXBK = block_size, nYBK = block_size;

[0075] When the block is on the right boundary of the image, nXBK = cols - j;

[0076] When the block is on the lower boundary of the image, nYBK = rows - i;

[0077] If j+block_size≤cols, then nXBK=block_size; otherwise nXBK=cols-j;

[0078] If i + block_size ≤ rows, then nYBK = block_size; otherwise nYBK = rows - i;

[0079] Calculate the screening boundary value:

[0080] Calculate the upper boundary threshold threshold_boundary_y_min=k×block_size;

[0081] Calculate the lower boundary threshold threshold_boundary_y_max=(1-k)×nYBK;

[0082] Calculate the left boundary threshold threshold_boundary_x_min=k×block_size;

[0083] Calculate the right boundary threshold threshold_boundary_x_max=(1-k)×nXBK;

[0084] Let m represent the vertex number of the detection box, m = 1, 2, 3, 4. For the four vertex coordinates (x_m, y_m) of the current detection box, if any vertex coordinate does not satisfy the following relationship:

[0085] threshold_boundary_x_min <x_m<threshold_boundary_x_max

[0086] and

[0087] threshold_boundary_y_min <y_m<threshold_boundary_y_max

[0088] It is determined that the detection frame is close to the boundary of the current block area.

[0089] In this embodiment, when determining whether the detection frame is close to the boundary in step S3, the boundary filtering coefficient k is set to 0.1, and its actual sizes nXBK and nYBK are calculated based on the current block position (starting coordinates i, j) and the total image size (rows, cols). Then, four filtering boundary thresholds are calculated, and the four vertex coordinates of each detection frame of the current block are traversed. If any vertex falls outside the core area surrounded by these filtering thresholds, the detection frame is determined to be close to the boundary and subsequent supplementary detection is required.

[0090] In step S4 of this embodiment, the formula for converting the center pixel coordinates (center_x, center_y) of the detection frame into the global center pixel coordinates (global_center_x, global_center_y) relative to the entire remote sensing image is:

[0091] global_center_x=center_x+j;

[0092] global_center_y=center_y+i.

[0093] In step S5 of this embodiment, the boundary coordinates of the newly intercepted block area are determined in the following manner:

[0094] Calculate the left margin:

[0095] new_block_x_min=max{global_center_x-block_size÷2,0};

[0096] Calculate the upper bound:

[0097] new_block_y_min=max{global_center_y-block_size÷2,0};

[0098] Calculate the right boundary:

[0099] new_block_x_max=min{global_center_x+block_size÷2,cols};

[0100] Calculate the lower bound:

[0101] new_block_y_max=min{global_center_y+block_size÷2,rows}.

[0102] In this embodiment, based on the global center coordinates (global_center_x, global_center_y) obtained in step S4, an attempt is made to intercept a new block of size block_size×block_size (1024x1024) with this point as the center. The max and min functions are used to ensure that the new block does not exceed the valid range of the entire scene image. The four boundary coordinates of the new block are specifically calculated. After the new rectangular area is determined using this boundary, the trained target detection model is called to perform target detection on the new block area again to generate a supplementary detection result.

[0103] In step S6 of this embodiment, the output of the final target detection result includes associating and storing the category information, position coordinates and confidence of each target, merging the initial detection results and supplementary detection results of all blocks, and associating the category information, position coordinate information and confidence of the merged results in units of a single detection frame, and writing them into the specified output file line by line.

[0104] In this embodiment, the position coordinates include the four vertex coordinate values ​​of the target frame. The position coordinate information refers to the pixel coordinate values ​​of the four corner points of the target detection frame. The output format is, for example, category information followed by the four vertex coordinates x1, y1, x2, y2, x3, y3, x4, y4, and finally the confidence level.

[0105] In this embodiment, the target detection model is a detection model that outputs a rotated target frame. The trained YOLO11-OBB model is used, and its optimal weight file best.pt is loaded.

[0106] In this embodiment, the high-resolution remote sensing image is a satellite remote sensing image, and the large target includes a ship.

[0107] The above description is only the optimal solution embodiment of the present invention and is not intended to limit the present invention. Various modifications or substitutions made by those skilled in the art without departing from the essence and protection scope of the present invention should also be within the protection scope of the present invention.

Claims

1. A large-scale target processing method based on remote sensing image segmentation, characterized in that: It includes the following steps: S1: Read the high-resolution remote sensing image, obtain the information of the number of rows "rows" and the number of columns "cols", set the block size "block_size" and the overlap threshold "overlop_threshold", calculate the block stride "stride", and perform block processing on the remote sensing image according to the block size "block_size", the block stride "stride", the number of rows "rows", and the number of columns "cols"; S2: Traverse each block area generated in step S1, call the best weights of the trained object detection model to perform object detection on the current block area, and generate detection results including object detection boxes and their categories, positions, and confidences; S3: Traverse each detection box generated in the current block area in step S2, and determine whether the detection box is close to the boundary of the current block area; S4: For the detection box determined to be close to the boundary of the current block area in step S3, calculate the center pixel coordinates (center_x, center_y) of the detection box, and convert the center pixel coordinates to the global center pixel coordinates (global_center_x, global_center_y) relative to the entire remote sensing image; S5: Centered on the global center pixel coordinates (global_center_x, global_center_y) calculated in step S4, intercept a rectangular area with the same size as the initial block on the entire remote sensing image. If the new block area exceeds the image boundary, only keep the valid area; call the best weights of the trained object detection model to perform object detection on the new block area again, and generate supplementary detection results; S6: Merge the detection results of step S2 and the supplementary detection results of step S5, and output the final object detection results to the storage medium.

2. A large-scale target processing method based on remote sensing image segmentation according to claim 1, characterized in that: In step S1, the formula for calculating the block stride "stride" is: stride = block_size × (1 - overlop_threshold).

3. The large-scale target processing method based on remote sensing image segmentation according to claim 1 is characterized in that: Step S3 further includes: Obtain the actual width nXBK pixels and the actual height nYBK pixels of the current block area. Let j be the starting coordinate of the current block horizontally, i be the starting coordinate of the current block vertically, and k be the boundary screening coefficient. k is a preset constant and satisfies 0 < k < 1. Then: When the block is in the non-boundary area of the image, nXBK = block_size, nYBK = block_size; When the block is on the right boundary of the image, nXBK = cols - j; When the block is on the lower boundary of the image, nYBK = rows - i; If j + block_size ≤ cols, then nXBK = block_size; otherwise nXBK = cols - j; If i + block_size ≤ rows, then nYBK = block_size; otherwise nYBK = rows - i; Calculate the screening boundary value: Calculate the upper boundary threshold threshold_boundary_y_min = k × block_size; Calculate the lower boundary threshold threshold_boundary_y_max=(1-k)×nYBK; Calculate the left boundary threshold threshold_boundary_x_min=k×block_size; Calculate the right boundary threshold threshold_boundary_x_max=(1-k)×nXBK; Let m represent the vertex number of the detection box, m = 1, 2, 3, 4. For the four vertex coordinates (x_m, y_m) of the current detection box, if any vertex coordinate does not satisfy the following relationship: threshold_boundary_x_min <x_m<threshold_boundary_x_max and threshold_boundary_y_min <y_m<threshold_boundary_y_max It is determined that the detection frame is close to the boundary of the current block area.

4. The large-scale target processing method based on remote sensing image segmentation according to claim 3 is characterized in that: In step S4, the formula for converting the center pixel coordinates (center_x, center_y) of the detection frame into the global center pixel coordinates (global_center_x, global_center_y) relative to the entire remote sensing image is: global_center_x=center_x+j; global_center_y=center_y+i.

5. The large-scale target processing method based on remote sensing image segmentation according to claim 1 is characterized in that: In step S5, the boundary coordinates of the newly intercepted block area are determined in the following manner: Calculate the left boundary: new_block_x_min=max{global_center_x-block_size÷2,0}; Calculate the upper bound: new_block_y_min=max{global_center_y-block_size÷2,0}; Calculate the right boundary: new_block_x_max=min{global_center_x+block_size÷2,cols}; Calculate the lower bound: new_block_y_max=min{global_center_y+block_size÷2,rows}.

6. The large-scale target processing method based on remote sensing image segmentation according to claim 1, characterized in that: In step S6, the outputting of the final target detection result includes associating and storing the category information, location coordinates and confidence level of each target.

7. The large-scale target processing method based on remote sensing image segmentation according to claim 6 is characterized in that: The position coordinates include the coordinate values ​​of the four vertices of the target frame.

8. The large-scale target processing method based on remote sensing image segmentation according to claim 1 is characterized in that: The target detection model is a detection model that outputs a rotated target frame.

9. The large-scale target processing method based on remote sensing image segmentation according to claim 1, characterized in that: The high-resolution remote sensing image is a satellite remote sensing image, and the large target includes a ship.

Images