Intelligent interpretation method for large-scale SAR image based on gradient feature pre-screening

Abstract

This invention provides a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening. The method includes: performing extreme value truncation on the acquired SAR image to obtain a target image; statistically analyzing the pixel value histogram of the target image and fitting the histogram to obtain peak pixel values; stretching the target image based on the peak pixel values ​​to obtain a standardized image; dividing the standardized image into multiple image patches using a sliding window; determining the gradient features of each image patch and selecting valid image patches whose gradient features meet preset conditions; interpreting the valid image patches to obtain local interpretation results; and fusing all local interpretation results based on their spatial location to obtain the full-image interpretation result. This method significantly reduces the number of image patches requiring deep learning inference, thereby greatly improving interpretation efficiency. Furthermore, by dynamically stretching the target image based on peak pixel values, it can better adjust the image contrast.

Description

Technical Field

[0001] This invention relates to the field of SAR image processing technology, and in particular to a method for intelligent interpretation of large-scale SAR images based on gradient feature pre-screening. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active remote sensing system operating in all weather conditions and around the clock, possesses the advantages of penetrating the earth's surface and high-resolution imaging, and is widely used in disaster monitoring and agricultural yield estimation. With the development of deep learning technology, intelligent interpretation of SAR images (such as target detection, classification, and segmentation) has shown great potential. However, SAR images are typically extremely large, with dimensions reaching tens of thousands of pixels. Direct full-image inference would incur extremely high computational costs and long processing times, and may not be compatible with the input size of deep learning models. Therefore, sliding window inference has become the mainstream solution for large-scale SAR images. This involves sliding a fixed-size window across the image, performing independent deep learning model inference on each image patch, and then stitching the results together to form a full-image interpretation.

[0003] However, SAR images contain a large number of areas lacking effective targets (such as grasslands and lakes). These areas appear as noise in the images, and deep learning inference on them would lead to a lot of redundant computation, wasting computing resources and reducing overall efficiency. Summary of the Invention

[0004] This invention provides an intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening, which solves the technical problems of low interpretation efficiency and serious waste of computing resources in existing large-scale SAR image interpretation technologies.

[0005] On the one hand, this invention provides a method for intelligent interpretation of large-scale SAR images based on gradient feature pre-screening, comprising:

[0006] The acquired SAR image is truncated to obtain the target image;

[0007] The pixel value histogram of the target image is statistically analyzed, and the pixel value histogram is fitted to obtain the peak pixel value;

[0008] Based on the peak pixel values, the target image is stretched to obtain a standardized image;

[0009] The standardized image is divided into multiple image blocks by a sliding window.

[0010] Determine the gradient features of each image block, and filter out valid image blocks whose gradient features meet preset conditions;

[0011] Each of the effective image blocks is interpreted to obtain the local interpretation results;

[0012] Based on the spatial location of each local interpretation result, all local interpretation results are merged to obtain the final full-map interpretation result.

[0013] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein the extreme value truncation processing of the acquired SAR image to obtain the target image includes:

[0014] Determine the truncation ratio of the SAR image;

[0015] Based on the truncation ratio, calculate the lower truncation threshold and the upper truncation threshold of the SAR image;

[0016] The pixel values ​​in the SAR image that are below the lower truncation threshold are adjusted to the lower truncation threshold, and the pixel values ​​that are above the upper truncation threshold are adjusted to the upper truncation threshold, thus obtaining the truncated target image.

[0017] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening includes: statistically analyzing the pixel value histogram of the target image and fitting the pixel value histogram to obtain peak pixel values, comprising:

[0018] Statistically analyze the distribution of all pixel values ​​in the target image;

[0019] The frequency of each pixel value in the target image is counted to obtain frequency statistics results.

[0020] Based on the frequency statistics and distribution, a pixel value histogram is generated;

[0021] Curve fitting is performed on the histograms of each pixel value to obtain the fitted histogram;

[0022] The pixel value that appears most frequently in the fitted histogram is used as the peak pixel value.

[0023] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein stretching the target image according to the peak pixel value to obtain a standardized image includes:

[0024] The peak pixel value is multiplied by a preset peak value to obtain the dynamic segmentation threshold;

[0025] The portion of the target image whose pixel value is less than or equal to the dynamic segmentation threshold is mapped to a first interval; wherein, the first interval is [0, preset intermediate segmentation value];

[0026] The portion of the target image whose pixel value is greater than the dynamic segmentation threshold is mapped to the second interval to obtain a standardized image; wherein, the second interval is (preset intermediate segmentation value, 255).

[0027] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein the standardized image is divided into multiple image patches by sliding window partitioning, including:

[0028] Determine the window width, window height, horizontal step size, and vertical step size of the sliding window;

[0029] Based on the total image width of the standardized image, the window width, and the horizontal step size, the maximum number of columns is obtained and used as the last column;

[0030] Based on the total image height of the standardized image, the window height, and the vertical step size, the maximum number of rows is obtained and used as the last row.

[0031] Starting from the top left corner of the standardized image, the sliding window is moved sequentially in the horizontal and vertical directions according to the horizontal step size and the vertical step size;

[0032] When the sliding window slides to the last column, the rightmost edge of the standardized image is used as the right edge of the sliding window to obtain the starting horizontal position of the sliding window;

[0033] When the sliding window slides to the last row, the bottom edge of the standardized image is used as the bottom edge of the sliding window to obtain the starting vertical position of the sliding window;

[0034] Extract the image region corresponding to each sliding window to generate multiple image blocks.

[0035] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein determining the gradient features of each image patch and screening out valid image patches whose gradient features satisfy preset conditions includes:

[0036] For each pixel in the image block, determine its neighborhood region and divide the neighborhood region into left and right parts and top and bottom parts;

[0037] Calculate the average brightness difference between the left and right parts and the average brightness difference between the top and bottom parts respectively;

[0038] Gradient features are obtained based on the average brightness difference between the left and right parts and the average brightness difference between the top and bottom parts.

[0039] If the gradient feature is greater than the preset intensity threshold, then the pixel is marked as a valid edge point;

[0040] Count the total number of valid edge points in an image patch;

[0041] If the total number is greater than a preset threshold, then the image block is marked as a valid image block.

[0042] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein the interpretation of each effective image block to obtain local interpretation results includes:

[0043] The effective image patch is input into a pre-trained deep learning model;

[0044] The deep learning model is used to interpret the input valid image blocks to obtain the corresponding local interpretation results;

[0045] The local interpretation results include at least one of bounding box coordinates, category labels, and confidence scores.

[0046] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein the step of fusing all local interpretation results according to the spatial location of each local interpretation result to obtain the final full-image interpretation result includes:

[0047] The local interpretation results are restored to their respective valid image blocks in the spatial position of the standardized image.

[0048] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening is provided, wherein restoring each local interpretation result to the spatial location of its respective effective image patch in the standardized image includes:

[0049] For the target detection results in the local interpretation results, the bounding box coordinates in each sliding window are superimposed with the offset of the corresponding sliding window, converted to the standardized image, and non-maximum suppression is applied to the detection boxes in the overlapping areas.

[0050] According to the present invention, a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening, wherein the step of fusing all local interpretation results according to the spatial location of each local interpretation result to obtain the final full-image interpretation result further includes:

[0051] For the pixel-level results in the local interpretation results, they are stitched to the standardized image according to the sliding window position, and pixel-level voting or weighted average processing is applied to the overlapping areas.

[0052] This invention provides a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening. Extreme value truncation removes extremely high / low pixel values ​​(such as radar noise or anomalous scattering) from the image, optimizing image processing and reducing noise interference with the interpretation results. Gradient feature pre-screening quickly identifies and skips areas lacking effective targets (such as grasslands, lakes, and deserts), significantly reducing the number of image patches requiring deep learning inference and thus greatly improving interpretation efficiency. Sliding window partitioning and pre-screening avoid redundant calculations on invalid regions, saving computational resources and time. Dynamic stretching of the target image based on peak pixel values ​​better adjusts image contrast, making the pixel value distribution more uniform, which helps enhance the contrast of effective targets. This method is applicable to different types of SAR images and has strong versatility. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening provided in this embodiment of the invention.

[0055] Figure 2 This is a schematic diagram of a large-scale SAR image sliding window provided in an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of local statistical gradient calculation provided in an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the architecture of the deep learning model for object detection provided in an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the structure of the large-scale SAR image intelligent interpretation system based on gradient feature pre-screening provided in an embodiment of the present invention;

[0059] Figure 6 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0061] Figure 1 This is a flowchart illustrating the intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening provided in this embodiment of the invention.

[0062] See Figure 1 A large-scale SAR image intelligent interpretation method based on gradient feature pre-screening may include the following steps 101 to 107.

[0063] Step 101: Perform extreme value truncation on the acquired SAR image to obtain the target image.

[0064] Specifically, step 101 may include the following steps:

[0065] Step 1: Determine the truncation ratio of the SAR image;

[0066] In this step, SAR images may contain extremely high / low pixel values ​​(such as radar noise or anomalous scattering), which need to be truncated to eliminate their impact on subsequent analysis. Specifically, a fixed truncation ratio can be selected, for example, 1%, 2%, or 5%, etc. This selection is usually based on understanding of SAR image data and accumulated experience. For images with high noise levels, a larger truncation ratio (such as 5%) can be selected; for images with low noise levels, a smaller truncation ratio (such as 1%) can be selected.

[0067] Step 2: Calculate the lower and upper truncation thresholds of the SAR image based on the truncation ratio.

[0068] In this step, the lower truncation threshold refers to the pixel value corresponding to the lowest truncation percentage in the image. For example, if a 2% truncation percentage is selected, the lower truncation threshold is the maximum value among the lowest 2% of pixel values ​​in the image. The upper truncation threshold refers to the pixel value corresponding to the highest truncation percentage in the image. For example, if a 2% truncation percentage is selected, the upper truncation threshold is the minimum value among the highest 2% of pixel values ​​in the image.

[0069] Step 3: Adjust the pixel values ​​in the SAR image that are below the lower truncation threshold to the lower truncation threshold, and adjust the pixel values ​​that are above the upper truncation threshold to the upper truncation threshold to obtain the truncated target image.

[0070] The above steps, by determining the truncation ratio, calculating the upper and lower truncation thresholds, and adjusting the pixel values, limit the pixel values ​​in the image to a reasonable range, which can effectively remove extreme noise in the image, optimize image processing effects, and improve interpretation efficiency and accuracy.

[0071] Step 102: Calculate the pixel value histogram of the target image and fit the pixel value histogram to obtain the peak pixel value.

[0072] Specifically, step 102 may include the following steps:

[0073] Step 1: Analyze the distribution of all pixel values ​​in the target image;

[0074] In this step, we can understand the overall distribution range of pixel values ​​in the target image, that is, the distribution from the minimum value to the maximum value.

[0075] Step 2: Count the frequency of each pixel value in the target image and obtain the frequency statistics results;

[0076] In this step, the frequency of each pixel value can be counted. For example, if a pixel with a value of 100 appears 1000 times in the image, this information will be recorded in the frequency statistics.

[0077] Step 3: Generate a pixel value histogram based on the frequency statistics and distribution.

[0078] In this step, a pixel value histogram is a graphical representation method. The horizontal axis represents the range of pixel values, and the vertical axis represents the frequency of each pixel value. Using the frequency statistics, the frequency of each pixel value can be plotted on a histogram, thus visually displaying the distribution of pixel values.

[0079] Step 4: Perform curve fitting on the histogram of each pixel value to obtain the fitted histogram;

[0080] In this step, since the histogram is composed of discrete frequency data, there may be some noise or irregular fluctuations. Curve fitting can fit these discrete data points into a smooth curve, thus more accurately reflecting the distribution pattern of pixel values. Various mathematical methods can be used for fitting, such as polynomial fitting and Gaussian fitting.

[0081] Step 5: Find the pixel value that appears most frequently in the fitted histogram and use it as the peak pixel value;

[0082] In this step, the peak pixel value refers to the pixel value with the highest frequency (vertical axis) in the fitted histogram.

[0083] The above steps eliminate noise fluctuations in the histogram through fitting, avoiding misjudgments of "false peaks" caused by SAR image noise. The peak pixel value serves as the threshold for subsequent dynamic stretching, ensuring that the stretching process adapts to the image content and enhances the contrast of effective targets.

[0084] Step 103: Stretch the target image based on the peak pixel value to obtain a standardized image.

[0085] Specifically, step 103 may include:

[0086] Step 1: Multiply the peak pixel value by a preset peak value to obtain the dynamic segmentation threshold;

[0087] Step 2: Map the pixels in the target image whose pixel values ​​are less than or equal to the dynamic segmentation threshold to the first interval; where the first interval is [0, preset intermediate segmentation value];

[0088] In this step, the pixel values ​​of the darker parts of the image are linearly stretched so that they are distributed within the first interval.

[0089] Step 3: Map the portion of the target image whose pixel value is greater than the dynamic segmentation threshold to the second interval to obtain a standardized image; wherein, the second interval is (preset intermediate segmentation value, 255];

[0090] In this step, the pixel values ​​of the brighter parts of the image are linearly stretched so that they are distributed within the second interval.

[0091] The above steps, through a two-stage stretching method, can better adjust the image contrast, making the pixel value distribution of the image more uniform and facilitating subsequent processing.

[0092] Typically, SAR images are stored in int16 data format. The data format of a standardized image can be converted to uint8 data format and then stored. Each pixel in int16 occupies 2 bytes, while uint8 only requires 1 byte. For large-scale SAR images (such as tens of thousands of pixels), this can significantly save storage costs. uint8 data consumes less memory and bandwidth, improving processing efficiency.

[0093] Specifically, the standardized image conforms to the following formula (1):

[0094] (1);

[0095] Wherein, PV is the original pixel value, that is, the original pixel value in the SAR image before processing; T is the pixel value corresponding to the truncation ratio, that is, the truncation threshold calculated according to the truncation ratio (e.g., 2%) in extreme value truncation processing; mP is the dynamic segmentation threshold, where m is the preset peak multiple and P is the peak pixel value; Q is the preset intermediate segmentation value; and NV is the pixel value in the normalized image, that is, the stretched pixel value.

[0096] Step 104: Divide the standardized image into multiple image blocks using a sliding window.

[0097] Specifically, step 104 may include the following steps:

[0098] Step 1: Determine the window width, window height, horizontal step, and vertical step of the sliding window;

[0099] In this step, the range of the sliding window can be represented by the following formula (2):

[0100] (2);

[0101] in, This is the starting point of the sliding window in the horizontal direction (it can also be understood as the starting point in the x-axis direction); This is the horizontal endpoint of the sliding window (or the endpoint in the x-axis direction). This is the starting point in the vertical direction of the sliding window (it can also be understood as the starting point in the y-axis direction); This is the endpoint in the vertical direction of the sliding window (or the endpoint in the y-axis direction). For window width, The height of the window; For horizontal step size, is the vertical step size. W is the width of the normalized image, and H is the height of the normalized image. For the maximum number of columns, This represents the maximum number of rows. For column numbers, This represents the number of rows. The formula above can be found in [reference needed]. Figure 2 , Figure 2 The arrows in the diagram indicate the four directions: up, down, left, and right, to make the plan clearer.

[0102] Step 2: Based on the total image width, window width, and horizontal step size of the standardized image, obtain the maximum number of columns, which will be used as the last column;

[0103] In this step, the difference between the total image width and the window width can be calculated, then the ratio of this difference to the horizontal step size can be calculated, and then 1 can be added to the ratio to obtain the maximum number of columns. This can be expressed by the following formula (3):

[0104] (3);

[0105] Step 3: Based on the total image height, window height, and vertical stride of the standardized image, obtain the maximum number of rows, which will be used as the last row;

[0106] In this step, the difference between the total image height and the window height can be calculated, then the ratio of this difference to the vertical step size can be calculated, and then 1 can be added to the ratio to obtain the maximum number of rows. This can be expressed by the following formula (4):

[0107] (4).

[0108] Step 4: Starting from the top left corner of the standardized image, move the sliding window sequentially in the horizontal and vertical directions according to the horizontal and vertical step sizes.

[0109] Step 5: When the sliding window slides to the last column, take the rightmost edge of the standardized image as the right edge of the sliding window to obtain the starting horizontal position of the sliding window;

[0110] In this step, it can be understood that when the slide window slides to the last column, the rightmost edge of the normalized image generally does not coincide with the right edge of the slide window, but is included within the slide window. That is, the rightmost edge of the normalized image is located to the left of the right edge of the slide window. In this case, by taking the rightmost edge of the normalized image as the right edge of the slide window and combining it with the window width of the slide window, the starting horizontal position of the slide window can be obtained.

[0111] Step 6: When the sliding window slides to the last row, take the bottom edge of the normalized image as the bottom edge of the sliding window to obtain the starting vertical position of the sliding window;

[0112] In this step, it can be understood that when the slide window slides to the last column, the bottom edge of the normalized image generally does not coincide with the bottom edge of the slide window, but is included within the slide window. That is, the bottom edge of the normalized image is located above the bottom edge of the slide window. In this case, by taking the bottom edge of the normalized image as the bottom edge of the slide window and combining it with the window height of the slide window, the starting vertical position of the slide window can be obtained.

[0113] Step 7: Extract the image region corresponding to each sliding window to generate multiple image blocks.

[0114] Furthermore, when the width of the last column of the image is less than the width of the sliding window, and the height of the last row of the image is less than the height of the sliding window, the traditional boundary handling method is as follows: instead of using the rightmost edge of the normalized image as the right edge of the sliding window, the right edge of the sliding window extends beyond the rightmost edge of the image, that is, it is located to the right of the right edge of the image; similarly, instead of using the bottom edge of the normalized image as the bottom edge of the sliding window, the bottom edge of the sliding window extends beyond the bottom edge of the image, that is, it is located below the bottom edge of the image. If the window extends beyond the image boundaries, invalid or zero-value data outside the boundaries may be introduced, which may lead to errors or biases in the analysis results, possibly requiring special boundary handling.

[0115] In the steps described above in this embodiment, the sliding window partitioning method allows for image block processing, enabling parallel processing of each image block and significantly improving overall computational efficiency. When processing image edges, aligning the right and bottom edges of the sliding window with the rightmost and bottommost edges of the image ensures that image edge areas are effectively processed, avoiding information loss. This solution, through edge alignment, eliminates the need for special processing or padding of window portions exceeding image boundaries, avoiding the shortcomings of traditional processing methods.

[0116] Step 105: Determine the gradient features of each image block and filter out valid image blocks whose gradient features meet the preset conditions.

[0117] Specifically, step 105 may include the following steps:

[0118] Step 1: For each pixel in the image block, determine its neighborhood region and divide the neighborhood region into left and right parts and top and bottom parts;

[0119] In this step, the neighborhood region (gradient feature window radius) can be predetermined. The gradient feature window is a square window, meaning half of its length is... ), preset intensity threshold And a preset number threshold (i.e., the minimum effective number of pixels threshold). In this embodiment, the window radius can be set as follows: The preset intensity threshold can be The preset quantity threshold can be This means that at least 25 pixels within the sliding window have a local statistical gradient greater than 1. Only when the weighted average values ​​on both sides of the gradient differ by more than two times can it be considered that there is a valid target within the sliding window.

[0120] Step 2: Calculate the average brightness difference between the left and right parts and the average brightness difference between the top and bottom parts respectively;

[0121] In this step, for example, starting from the top left pixel of the sliding window, the left and right sides of the current pixel are calculated one by one. The weighted average of pixels within the (height × width) region is and upper and lower sides The weighted average of pixels within the region is and The weighted average is shown in formula (5) below:

[0122] (5);

[0123] in, The coordinates of the current pixel. The distance to the current point's coordinates; negative values ​​are used to move left / up, and positive values ​​are used to move right / down. The pixel values ​​of the neighboring points in the image that are used in the mean calculation. The weights can be set to equal weighting, distance-related weighting, etc. This embodiment uses equal weighting, i.e. The above formula is as follows: Figure 3 As shown.

[0124] Average brightness difference between the left and right parts As shown in formula (6):

[0125] (6);

[0126] Average brightness difference between the upper and lower parts As shown in formula (7):

[0127] (7);

[0128] Step 3: Based on the average brightness difference between the left and right parts and the average brightness difference between the top and bottom parts, obtain the gradient features;

[0129] In this step, the statistical gradient of the vertical region is calculated. and the statistical gradient of the horizontal region As shown in formula (8):

[0130] (8);

[0131] The gradient feature G is shown in the following formula (9):

[0132] (9).

[0133] Using the above method, the regional statistical gradient of each pixel within the entire sliding window range can be calculated from left to right and from top to bottom.

[0134] First use and The weighted average of all pixels within the sliding window range is calculated using a window of a certain size. Then, the average value of the corresponding region is directly taken into the calculation, which can reduce redundant calculations and improve processing efficiency.

[0135] Step 4: If the gradient feature is greater than the preset intensity threshold, then mark the pixel as a valid edge point;

[0136] Step 5: Count the total number of valid edge points in the image patch;

[0137] Step 6: If the total number is greater than the preset number threshold, then mark the image block as a valid image block.

[0138] SAR images contain a large amount of speckle noise, caused by the interference between radar echo signals and complex scattering bodies of ground objects. Directly using gradient calculations will produce a large number of "pseudo" high gradient values, which cannot be used as an indicator of changes in ground features. Therefore, a local statistical gradient is adopted as the gradient calculation strategy to reduce the impact of speckle noise. If the total number is less than or equal to a preset threshold, the marked image patch is ignored.

[0139] Step 106: Interpret each valid image block to obtain the local interpretation results.

[0140] Specifically, step 106 may include the following steps:

[0141] Step 1: Input the valid image patches into the pre-trained deep learning model;

[0142] Step 2: Using a deep learning model, interpret the valid input image blocks to obtain the corresponding local interpretation results;

[0143] The local interpretation results may include at least one of the following: bounding box coordinates, category labels, and confidence scores.

[0144] In the above steps, a valid image patch contains the region of the potential target. Bounding box coordinates, for object detection tasks, the model outputs the target's position within the image patch, usually represented by a bounding box, including the coordinates of the top-left and bottom-right corners. Category label, the model determines the category of the target in the image patch, such as "oil tank," "building," or "vehicle." Confidence score, the model provides a confidence score for each interpretation result, representing how confident the model is of that result; it is typically a value between 0 and 1. The deep learning model can be of type Faster R-CNN, U-Net, etc., and the model type can be selected according to the task requirements (e.g., object detection, semantic segmentation).

[0145] The aforementioned deep learning model can be used for various types of SAR image interpretation tasks, such as target detection, semantic segmentation, instance segmentation, and change detection. Taking target detection as an example, the model structure uses the ConvNeXt Base network as the feature extraction backbone, FPN as the feature enhancement module, and Faster RCNN as the bounding box regression and classification head to extract intelligent interpretation results of oil tank targets. The output results should include the bounding box coordinates (local coordinates) of the target, the target category, and the target confidence score.

[0146] By leveraging deep learning models for inference, automated and intelligent interpretation of effective image patches can be achieved, thereby improving interpretation efficiency and accuracy. Since local interpretation results can include various types of information, this method is applicable to a wide range of SAR image interpretation tasks, such as target detection and semantic segmentation, demonstrating flexibility and versatility.

[0147] Examples of deep learning model structures can be found as follows: Figure 4 As shown. Figure 4 The diagram illustrates the structure of a Convolutional Neural Network (CNN). The leftmost layer represents the input layer, where the input is a SAR image. The input SAR image first passes through a series of convolutional layers, progressively extracting image features, and then the feature maps are fed into a pyramid pooling module. Pyramid pooling aggregates features from different scales, thereby capturing multi-scale information in the image. The output of the pyramid pooling module is fed into a fusion layer, which may fuse feature maps from different scales to retain more spatial information. A 4x upsampling operation is used to increase the resolution of the feature maps to match the resolution of the original input image.

[0148] The upsampled feature maps are fed into the RPN, which generates a series of candidate regions (proposals) that may contain the target. The RPN typically consists of two branches: one for classification (determining whether a candidate region contains the target) and the other for bounding box regression (adjusting the position and size of the candidate regions). The candidate regions output by the RPN are further processed to generate the final proposals. These proposals are regions that the network believes may contain the target, and they are fed into subsequent network layers for classification and bounding box refinement. Finally, these proposals are fed into a classifier and a bounding box regressor. The classifier determines the category of the target in each proposal, while the bounding box regressor further refines the position and size of each proposal.

[0149] Step 107: Based on the spatial location of each local interpretation result, merge all local interpretation results to obtain the final full-map interpretation result.

[0150] In this embodiment, extreme value truncation can remove extremely high / low pixel values ​​(such as radar noise or anomalous scattering) from the image, optimizing image processing and reducing noise interference with the interpretation results. Gradient feature pre-screening can quickly identify and skip areas lacking effective targets (such as grasslands, lakes, deserts, etc.), significantly reducing the number of image patches requiring deep learning inference and thus greatly improving interpretation efficiency. Sliding window partitioning and pre-screening avoid redundant calculations of invalid regions, saving computational resources and time. Dynamic stretching of the target image based on peak pixel values ​​allows for better adjustment of image contrast, resulting in a more uniform pixel value distribution, which helps enhance the contrast of effective targets. This embodiment is applicable to different types of SAR images and has strong versatility.

[0151] In one embodiment of this specification, based on the spatial location of each local interpretation result, all local interpretation results are fused to obtain the final full-map interpretation result, which may include:

[0152] The local interpretation results are restored to the spatial location of their respective valid image patches in the standardized image.

[0153] In one embodiment of this specification, restoring the local interpretation results to the spatial location of their respective valid image patches in the standardized image may include:

[0154] For target detection results in the local interpretation results, the bounding box coordinates within each sliding window are superimposed with the corresponding sliding window offset, converted to a standardized image, and non-maximum suppression is applied to the detection boxes in the overlapping areas.

[0155] In this embodiment, during the sliding window inference process, the target detection result within each sliding window is represented in local coordinates (i.e., coordinates relative to the top-left corner of the sliding window). These local coordinates need to be transformed to the global coordinate system of the normalized image. Specifically, the bounding box coordinates within each sliding window are added to the corresponding sliding window's offset (i.e., the coordinates of the top-left corner of the sliding window in the normalized image).

[0156] Because sliding windows may overlap, the same target may be detected multiple times. To eliminate this redundant detection, non-maximum suppression (NMS) is used. By comparing the confidence of the detection boxes, the detection box with the highest confidence is retained, while other detection boxes with high overlap are suppressed. This ensures that each target is detected only once, avoiding duplicate detection.

[0157] In this embodiment, by converting local coordinates to global coordinates, it is ensured that all detection results can be correctly restored to their original image positions, providing accurate positioning information for subsequent global analysis and applications. Non-maximum suppression can effectively reduce the problem of multiple detections caused by sliding window overlap, improving the accuracy and reliability of the interpretation results.

[0158] In one embodiment of this specification, based on the spatial location of each local interpretation result, all local interpretation results are fused to obtain the final full-map interpretation result, which may further include:

[0159] For pixel-level results in local interpretation, they are stitched to a standardized image according to the sliding window position, and pixel-level voting or weighted averaging is used for overlapping areas.

[0160] In this embodiment, for pixel-level interpretation results (such as semantic segmentation or instance segmentation results), the interpretation results within each sliding window need to be stitched into the normalized image according to the position of the sliding window. This means placing the pixel-level prediction result within each sliding window into its corresponding position in the normalized image. Since there is overlap between sliding windows, the prediction results of multiple sliding windows will cover the same area in the stitched image.

[0161] Pixel-wise voting involves counting the number of votes for each category in multiple sliding window predictions for each pixel in the overlapping region, and selecting the category with the most votes as the final result. This method is simple and intuitive, and suitable for situations with a small number of categories. Weighted averaging, on the other hand, calculates a weighted average of multiple sliding window predictions for each pixel in the overlapping region. The weights can be determined based on the confidence level, location, or other factors of the sliding windows. This method can better utilize the information from multiple sliding windows, improving the smoothness and accuracy of the interpretation results.

[0162] Based on the same general inventive concept, this invention also protects a large-scale SAR image intelligent interpretation system based on gradient feature pre-screening, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of the structure of the intelligent interpretation system for large-scale SAR images based on gradient feature pre-screening provided in an embodiment of the present invention. The intelligent interpretation system for large-scale SAR images based on gradient feature pre-screening provided in this invention is described below. The intelligent interpretation system for large-scale SAR images based on gradient feature pre-screening described below can be referred to in correspondence with the intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening described above.

[0163] The large-scale SAR image intelligent interpretation system based on gradient feature pre-screening includes a truncation module 501, a fitting module 502, a stretching module 503, a sliding window module 504, a gradient module 505, an interpretation module 506, and a fusion module 507.

[0164] The truncation module 501 is used to perform extreme value truncation processing on the acquired SAR image to obtain the target image;

[0165] The fitting module 502 is used to statistically analyze the pixel value histogram of the target image and fit the pixel value histogram to obtain the peak pixel value;

[0166] The stretching module 503 is used to stretch the target image based on the peak pixel value to obtain a standardized image;

[0167] The sliding window module 504 is used to divide the standardized image into multiple image blocks using a sliding window.

[0168] The gradient module 505 is used to determine the gradient features of each image block and filter out valid image blocks whose gradient features meet preset conditions.

[0169] The interpretation module 506 is used to interpret each valid image block to obtain the local interpretation results;

[0170] The fusion module 507 is used to fuse all local interpretation results according to the spatial location of each local interpretation result to obtain the final full-map interpretation result.

[0171] Figure 6 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0172] like Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions from the memory 630 to execute a large-scale SAR image intelligent interpretation method based on gradient feature pre-screening.

[0173] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0174] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the large-scale SAR image intelligent interpretation method based on gradient feature pre-screening provided by the above methods.

[0175] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening provided by the above methods.

[0176] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0177] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent interpretation of large-scale SAR images based on gradient feature pre-screening, characterized in that, include: The acquired SAR image is truncated to obtain the target image; The pixel value histogram of the target image is statistically analyzed, and the pixel value histogram is fitted to obtain the peak pixel value; Based on the peak pixel values, the target image is stretched to obtain a standardized image; The standardized image is divided into multiple image blocks by a sliding window. Determining the gradient features of each image block and filtering out valid image blocks whose gradient features meet preset conditions includes: for each pixel in an image block, determining its neighborhood region and dividing the neighborhood region into left and right parts and top and bottom parts; calculating the average brightness difference between the left and right parts and the average brightness difference between the top and bottom parts, specifically starting from the top left pixel of the sliding window, calculating the weighted average value of pixels in the left and right regions of the current pixel one by one. and The weighted average of pixels in the upper and lower regions and The difference in average brightness between the left and right parts As shown below: ； Average brightness difference between the upper and lower parts As shown below: ； Gradient features are obtained based on the average brightness difference between the left and right parts and the average brightness difference between the top and bottom parts. If the gradient feature is greater than a preset intensity threshold, then the pixel is marked as a valid edge point; count the total number of valid edge points in the image block; if the total number is greater than a preset number threshold, then the image block is marked as a valid image block; Each of the effective image blocks is interpreted to obtain the local interpretation results; Based on the spatial location of each local interpretation result, all local interpretation results are merged to obtain the final full-map interpretation result.

2. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 1, characterized in that, The process of performing extreme value truncation on the acquired SAR image to obtain the target image includes: Determine the truncation ratio of the SAR image; Based on the truncation ratio, calculate the lower truncation threshold and the upper truncation threshold of the SAR image; The pixel values ​​in the SAR image that are below the lower truncation threshold are adjusted to the lower truncation threshold, and the pixel values ​​that are above the upper truncation threshold are adjusted to the upper truncation threshold, thus obtaining the truncated target image.

3. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 1, characterized in that, The step of statistically analyzing the pixel value histogram of the target image and fitting the pixel value histogram to obtain the peak pixel value includes: Statistically analyze the distribution of all pixel values ​​in the target image; The frequency of each pixel value in the target image is counted to obtain frequency statistics results. Based on the frequency statistics and distribution, a pixel value histogram is generated; Curve fitting is performed on the histograms of each pixel value to obtain the fitted histogram; The pixel value that appears most frequently in the fitted histogram is used as the peak pixel value.

4. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 1, characterized in that, The step of stretching the target image based on the peak pixel value to obtain a standardized image includes: The peak pixel value is multiplied by a preset peak value to obtain the dynamic segmentation threshold; The portion of the target image whose pixel value is less than or equal to the dynamic segmentation threshold is mapped to a first interval; wherein, the first interval is [0, preset intermediate segmentation value]; The portion of the target image whose pixel value is greater than the dynamic segmentation threshold is mapped to the second interval to obtain a standardized image; wherein, the second interval is (preset intermediate segmentation value, 255).

5. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 1, characterized in that, The step of dividing the standardized image into multiple image blocks using a sliding window method includes: Determine the window width, window height, horizontal step size, and vertical step size of the sliding window; Based on the total image width of the standardized image, the window width, and the horizontal step size, the maximum number of columns is obtained and used as the last column; Based on the total image height of the standardized image, the window height, and the vertical step size, the maximum number of rows is obtained and used as the last row. Starting from the top left corner of the standardized image, the sliding window is moved sequentially in the horizontal and vertical directions according to the horizontal step size and the vertical step size; When the sliding window slides to the last column, the rightmost edge of the standardized image is used as the right edge of the sliding window to obtain the starting horizontal position of the sliding window; When the sliding window slides to the last row, the bottom edge of the standardized image is used as the bottom edge of the sliding window to obtain the starting vertical position of the sliding window; Extract the image region corresponding to each sliding window to generate multiple image blocks.

6. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 1, characterized in that, The process of interpreting each of the effective image blocks to obtain local interpretation results includes: The effective image patch is input into a pre-trained deep learning model; The deep learning model is used to interpret the input valid image blocks to obtain the corresponding local interpretation results; The local interpretation results include at least one of bounding box coordinates, category labels, and confidence scores.

7. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 1, characterized in that, The step of fusing all local interpretation results based on their spatial locations to obtain the final full-map interpretation result includes: The local interpretation results are restored to their respective valid image blocks in the spatial position of the standardized image.

8. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 7, characterized in that, The step of restoring each of the local interpretation results to the spatial location of its respective valid image block in the standardized image includes: For the target detection results in the local interpretation results, the bounding box coordinates in each sliding window are superimposed with the offset of the corresponding sliding window, converted to the standardized image, and non-maximum suppression is applied to the detection boxes in the overlapping areas.

9. The intelligent interpretation method for large-scale SAR images based on gradient feature pre-screening according to claim 7, characterized in that, The step of fusing all local interpretation results based on their spatial locations to obtain the final full-map interpretation result also includes: For the pixel-level results in the local interpretation results, they are stitched to the standardized image according to the sliding window position, and pixel-level voting or weighted average processing is applied to the overlapping areas.

Images