Sensing, retrieving and positioning target detection method and system for synthetic aperture radar remote sensing image, storage medium and product

Abstract

The invention discloses a sensing, retrieval and positioning target detection method and system for a synthetic aperture radar remote sensing image, a storage medium and a product, belongs to the technical field of computer vision and remote sensing image processing, and solves the problem of insufficient utilization of group distribution characteristics caused by a single-target independent detection strategy. A target initial small group box is generated through a weak detector based on a synthetic aperture radar remote sensing image, and position correction is carried out; a middle group detection frame and a large group detection frame are generated step by step based on a merged group generation algorithm and a target initial small group frame before correction, and after size processing of each cutting area is completed through a Gaussian size adjustment module, the cut areas are respectively input into a middle group detector and a large group detector for detection. Obtaining a corrected middle group frame and a corrected large group frame; and inputting each corrected group box into a progressive detector based on a box integration fusion algorithm, and processing to obtain a predicted group box. The method is used for target detection in the synthetic aperture radar remote sensing image.

Description

Technical Field

[0001] A method, system, storage medium, and product for sensing, retrieving, and locating targets in synthetic aperture radar (SAR) remote sensing images are disclosed. This invention belongs to the field of computer vision and remote sensing image processing technology. Background Technology

[0002] Ship detection based on Synthetic Aperture Radar (SAR) remote sensing images has always been a hot topic in academic research in the field of Earth observation, and has important application value in transnational trade, environmental protection, fishery production and national defense security.

[0003] However, most current ship detection models for SAR remote sensing images employ an independent single-target detection strategy, typically treating each ship as an individual for identification and localization. These methods neglect the potential spatial relationships and group distribution characteristics between ships during the modeling process, making it difficult for the model to fully utilize information provided by the surrounding environment and neighboring targets. Especially in densely populated ports or nearshore areas, relying solely on single-target information often weakens the model's overall target perception ability, leading to insufficient representation of individual ship features and poor localization performance. Existing methods still suffer from the following prominent problems when detecting densely distributed ship groups: First, the single-target independent detection strategy leads to insufficient utilization of group distribution characteristics. Traditional CFAR algorithms and mainstream deep learning models such as Faster R-CNN, YOLO, and DETR all employ a single-target independent detection strategy, ignoring the group distribution characteristics of ships. This independent detection mechanism has certain shortcomings when facing densely distributed ship groups. In densely populated areas such as ports and waterways, ships often exhibit clustered distribution characteristics. However, existing methods detect each ship as an independent individual, ignoring the spatial correlation and group structure information between ships. When ships are densely arranged, existing methods struggle to accurately distinguish adjacent ships, easily leading to problems such as overlapping detection boxes and confused target attribution. In dense port environments, the spacing between ships may be small, and traditional horizontal bounding box detection models cannot accurately describe the shape and positional relationships of these targets, resulting in a large number of false positives and false negatives.

[0004] Second, existing methods are insufficient in perceiving spatial environmental information of single targets, especially in detecting densely distributed ship groups. Due to the lack of effective perception of the spatial environment surrounding the target, the detection accuracy of existing methods drops significantly in complex scenes. In near-shore areas, the problem of insufficient spatial perception is even more severe due to the extremely complex background environment. The coastline causes serious interference; ships may visually resemble background clutter, making them difficult to distinguish from surrounding buildings and other elements. In some cases, ships may even overlap with background features, further increasing the detection difficulty. Existing methods are more focused on detecting large ships in the open ocean, while the detection performance of near-shore ships is severely limited by the interference effect of the coastline.

[0005] Third, existing methods suffer from poor data correlation in densely distributed ship swarm scenarios, leading to a significant decrease in detection accuracy and positioning precision. In dense ship environments, determining data correlation becomes challenging; ships detected in remote sensing images may be incorrectly associated with other observations, frequently resulting in erroneous or inaccurate maritime monitoring findings. This erroneous association is particularly common when ships are densely packed, as adjacent ships share high feature similarities, making accurate differentiation difficult with existing methods. In high traffic density areas, frequent ship occlusion and overlap further blur target boundaries, increasing the risk of missed or false detections and reducing the reliability and robustness of detection models in practical deployments. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, storage medium, and product for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images, addressing the problems of insufficient utilization of group distribution characteristics due to independent single-target detection strategies; inadequacies in sensing spatial environmental information of single targets; and poor data correlation in densely distributed ship swarm scenarios, leading to a significant decrease in detection accuracy and positioning accuracy.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images includes the following steps: Step 1: Acquire the synthetic aperture radar remote sensing image to be detected; Step 2: Input the synthetic aperture radar remote sensing image into the region candidate network of Faster R-CNN to generate the initial small group box of the target, and use a convolutional neural network to correct the position of the initial small group box of the target to obtain the corrected small group box. The convolutional neural network includes a Rol extractor and a classification network connected in sequence. Step 3: Based on the merged group generation algorithm and the initial small group bounding box of the target before correction, generate medium group detection boxes and large group detection boxes step by step. Then, crop out the regions of the synthetic aperture radar remote sensing image corresponding to the medium group detection boxes and large group detection boxes respectively. After the size processing of each cropped region is completed by the Gaussian size adjustment module, the regions are input into the medium group detector and the large group detector respectively for detection to obtain the corrected medium group bounding box and large group bounding box. Step 4: Construct a progressive detector for perception-retrieval-localization, including a guided feature sampling module and a guided bounding box iterative optimization module; Step 5: Input the corrected group boxes into the progressive detector to obtain all detection boxes, and sort them in descending order of confidence based on the box fusion algorithm. Select the box with the highest confidence as the main box, traverse the remaining boxes, and compare the intersection-union ratio with the main box. If it is greater than the threshold, perform weighted fusion; otherwise, keep it as a new box. Finally, the predicted group boxes are obtained.

[0008] Furthermore, during the training phase, the predicted group boxes from steps 2-5 and the ground truth bounding boxes are modeled as two-dimensional Gaussian distributions, with the center coordinates of the boundary as the mean vector of the Gaussian distribution and the covariance matrix constructed from the squares of half the width and height of the bounding box. The training is performed by calculating the second-order Wasserstein distance loss function between the two two-dimensional Gaussian distributions.

[0009] Furthermore, in step 3, the specific steps for generating the middle group detection box and the large group detection box step by step based on the merged group generation algorithm and the initial small group box of the target before correction are as follows: Step 2.1: Receive the set of initial small group boxes of the target before correction and the confidence level of each initial small group box of the target before correction; Step 2.2: Select the initial small group box of the target before correction with the highest confidence level as the merging center; Step 2.3: Traverse the remaining initial small group boxes of the target before correction and the intersection-union ratio of the merge center. , will be compared Greater than a given threshold All the initial small group boxes of the target before correction are merged into a medium group detection box; Step 2.4: If the number of detection boxes in the group reaches... End the merging process and discard the unmerged target initial small group boxes before correction to obtain the final multiple medium group detection boxes. Otherwise, select the target initial small group box before correction with the highest confidence from the unmerged medium group boxes as the merging center and execute step 2.4 again to merge the current unmerged target initial small group boxes before correction. Step 2.5: Sort the medium-group detection boxes according to their confidence scores and take the medium-group detection box with the highest confidence score as the benchmark. Traverse the remaining medium-group detection boxes and merge the medium-group detection boxes whose intersection-union ratio with the benchmark exceeds a given threshold. If the merged region covers the outer range of all medium-group detection boxes, a large-group detection box is obtained. Otherwise, traverse the remaining medium-group detection boxes and continue merging.

[0010] Furthermore, the specific implementation steps of the Gaussian size adjustment module in step 3 are as follows: Step 3.1: Calculate the absolute size of all targets in each clipping region cut out by the middle group detection box and the large group detection box, and calculate the average absolute size based on the absolute size of all targets; The formula for the absolute size of all targets is: in, For the first Within the first cropping area One goal width, For the first Within the first cropping area One goal height, Indicate target The absolute size; Step 3.2: Randomly sample the corresponding scaling factor from a Gaussian distribution with the mean absolute size; Step 3.3: Scale the corresponding cropping area according to the scaling factor; Step 3.4: Determine whether each scaled cropped area needs to be filled or cropped. If the scaled cropped area is smaller than the given size, fill it; if the scaled cropped area is larger than the given size, crop it.

[0011] Furthermore, the specific implementation steps of the guided feature sampling module are as follows: Step 4.1: Flatten the small group boxes after position correction and the medium and large group boxes after scale correction. Based on the flattening results, generate the query vector and key vector required by the attention mechanism through two linear layers respectively. Step 4.2: The nonlocal neural network obtains the corrected small group boxes, medium group boxes, and large group boxes from the input; Using three 1×1×1 convolution mapping functions , and Generate corresponding query features, key features, and value features respectively. Multiply the query features and key features of each group by the feature similarity weight, and then multiply the feature similarity weight by the value feature to obtain the aggregated features. Using an output mapping The aggregated features are linearly mapped, and the mapped features are added to the synthetic aperture radar remote sensing image to enhance the input synthetic aperture radar remote sensing image. The enhanced image is passed through three residual convolutional layers, then through a linear mapping layer for feature projection, and combined with the key vector to generate the value vector in the attention mechanism; Step 4.3: The query vectors corresponding to each group pass through a linear mapping layer, and then through three attention heads (Head 1, Head 2, and Head 3) corresponding to the corrected large group bounding boxes, medium group bounding boxes, and small group bounding boxes, respectively, to output a set of two-dimensional offsets. ,in, Indicates the first The first attention point in the The first query vector corresponding to the first At each sampling point Direction shift, Indicates the first The first attention point in the The first query vector corresponding to the first At each sampling point The direction is shifted, and at the same time, the query vectors corresponding to each group are normalized into a set of attention weights by the Softmax function after passing through another linear mapping layer. Step 4.4: Construct the feature spaces of the corrected small group boxes, medium group boxes, and large group boxes using the value vectors output in Step 4.2. Then, use three attention heads, Head 1, Head 2, and Head 3, to perform sparse sampling from the feature spaces of the corresponding large group boxes, medium group boxes, and small group boxes to obtain the corresponding salient features. Step 4.5: For each attention head, the corresponding salient features are sampled and aggregated with the learned attention weights to obtain aggregated sampled values. Then, a linear transformation is applied to all aggregated features to output an aggregated feature image.

[0012] Furthermore, the specific implementation steps of the guided frame iterative optimization module are as follows: Step 5.1: First, crop the corrected large group boxes, medium group boxes, and small group boxes, and then perform a flattening operation on the cropped feature maps to obtain a feature sequence with uniform feature dimensions. Step 5.2: Align the feature sequences of different groups in dimensions using a fully connected layer; Step 5.3: Concatenate the feature sequences of the aligned large group boxes, medium group boxes, and small group boxes to different layers of the Transformer decoder. Use three iterations to achieve the fusion of group information and decoding process, and finally output a single detection box. The bottom layer decoder incorporates the features corresponding to the large group box, the middle layer decoder incorporates the features corresponding to the medium group box, and the top layer decoder incorporates the features corresponding to the small group box.

[0013] Furthermore, in step 5.3, the specific steps for fusing group information with the decoding process using three iterations to finally output a single ship detection frame are as follows: First, the aggregated feature image output by the guided feature sampling module is used as the initial predicted bounding box. Under the spatial constraints of the corrected large group of bounding boxes, the first round of bounding box regression is performed to obtain the updated first round of predicted bounding boxes, as shown in the formula: in, This represents the predicted bounding box obtained in the first round. This represents the sigmoid activation function. This represents the inverse sigmoid function. This represents the regression offset in the first round, which is the two-dimensional offset of the large group bounding box. This refers to the reference box used before the first round of regression begins, i.e., the aggregated feature image; Then, using the predicted bounding boxes obtained in the first round as input, a second round of bounding box regression is performed under the corrected mid-group bounding box spatial constraints to obtain the updated predicted bounding boxes for the second round. The formula is as follows: in, This represents the predicted bounding box obtained in the second round. This represents the regression offset in the second round, which is the two-dimensional offset of the middle group box; Finally, based on the predicted bounding boxes from the second round, a third round of bounding box regression is performed under the spatial constraints of the corrected small group of bounding boxes to obtain the updated predicted bounding boxes for the third round. The formula is as follows: in, This represents the predicted bounding box obtained in the third round. This represents the regression offset in the third round, which is the two-dimensional offset of the small group bounding box; Through the above three iterations, the regression offsets of the corrected large group boxes, medium group boxes, and small group boxes are gradually corrected, achieving progressive and accurate localization of the detection boxes, and finally outputting the single ship detection box.

[0014] A target detection system for sensing, retrieving, and locating targets in synthetic aperture radar (SAR) remote sensing images includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of a method for sensing, retrieving, and locating targets in SAR remote sensing images.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images.

[0016] A computer program product includes a computer program that, when executed by a processor, implements the steps of a method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images.

[0017] Compared with the prior art, the advantages of the present invention are as follows: I. This invention introduces a group-level modeling and a guided progressive detection mechanism, which makes the detection process less sensitive to local occlusion, weak scattering and noise interference when facing densely distributed targets, significant scale differences and complex sea clutter / ground object scattering backgrounds. The detection results show higher consistency and stability in different scenarios, thereby reducing the risk of significant fluctuations in model performance (referring to steps 2-5) with changes in scenario. Second, this invention uses a weak detector (Faster R-CNN region candidate network) to generate initial small group boxes of targets, and combines a group merging strategy with guided feature sampling and iterative regression mechanism. Even if there is a certain deviation in the initial detection results, the detection box position can still be gradually corrected under the group space constraint. This effectively avoids the strong dependence of existing multi-stage detection methods on high-quality initial candidate boxes, and improves the applicability and fault tolerance of the method in low signal-to-noise ratio, weak scattering and complex nearshore environments. Third, this invention introduces explicit group space constraints during feature sampling (i.e., feature sampling by the guided feature sampling module) and bounding box iteration optimization (i.e. bounding box iteration optimization by the guided bounding box iteration optimization module), guiding the attention mechanism to perform feature retrieval within semantically meaningful regions. This avoids existing Transformer-type methods generating a large amount of invalid attention in complex remote sensing backgrounds, thereby improving the effectiveness of attention computation and resource utilization efficiency without significantly increasing model complexity, which is more conducive to practical engineering deployment and long-term operation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the three-level group structure in this invention; Figure 3 This is a schematic diagram of the framework structure of the present invention; Figure 4 This is a schematic diagram of the group merging generation algorithm in this invention; Figure 5 This is a schematic diagram of the Gaussian scale adjustment module in this invention; Figure 6 This is a schematic diagram of the guided feature sampling module structure in this invention; Figure 7 This is a schematic diagram of the guided frame iterative optimization module structure in this invention; Figure 8 This is a schematic diagram of the frame integration and fusion algorithm process in this invention; Figure 9 This is a schematic diagram of the structure of a computer device according to the present invention; Figure 10 This is a comparison of the accuracy results of experiments conducted on the SSDD and HRSID datasets between the present invention and existing technologies. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In one exemplary embodiment, this application provides a method for sensing, retrieving, and locating target detection in synthetic aperture radar remote sensing images, such as... Figure 1 As shown, the frame structure is as follows Figure 2 As shown, the specific steps are as follows: Step 1: Acquire the synthetic aperture radar remote sensing image to be detected; Experiments were conducted using the SSDD and HRSID datasets. The SSDD dataset includes 1160 SAR remote sensing images with a training set to test set ratio of 4:1. The HRSID dataset includes 5604 SAR remote sensing images with a training set to test set ratio of 13:7. All images are from multi-source SAR sensors such as Radarsat-2, TerraSAR-X, and Sentinel-1.

[0022] A large swarm (LS) refers to the largest union of the bounding boxes of all ships, which is the overall area where ships may appear; a medium swarm (MS) represents a sub-region formed by several ships; and a small swarm (SS) represents a single ship. The three-level swarm structure is as follows: Figure 3 As shown.

[0023] Step 2: Input the synthetic aperture radar remote sensing image into the region proposal network (RPN) of Faster R-CNN. Generate initial small group bounding boxes for the target, resulting in a set of initial small group bounding boxes, whose confidence level is denoted as . Its quantity is recorded as The initial small group bounding boxes of the target are corrected using a convolutional neural network to obtain corrected small group bounding boxes. The convolutional neural network includes a Rool extractor and a classification network connected in sequence. The Rool extractor cropped a 7×7 feature map, and the classification network decoupled the cropped 7×7 feature map to obtain better spatial location information and semantic information at the same time. The classification network is a fully connected layer for classification and regression, and a weak detector (region candidate network of Faster R-CNN) is used to pay attention to more loosely distributed group information and avoid missing some detection targets, so as to make the initial small group generation more accurate and facilitate the subsequent generation of medium and large groups.

[0024] Step 3: Based on the merged group generation algorithm and the initial small group bounding boxes of the target before correction, generate medium group and large group bounding boxes step by step to construct a three-level group structure (i.e., the corrected small group bounding boxes, medium group bounding boxes, and large group bounding boxes). Then, crop out the regions of the synthetic aperture radar remote sensing image corresponding to the medium group bounding boxes and large group bounding boxes respectively. After the size processing of each cropped region is completed by the Gaussian size adjustment module, it is input into the medium group detector and the large group detector respectively for detection to obtain the corrected medium group bounding boxes and large group bounding boxes. Among them, the medium group detector is used for local dense area detection, and the large group detector is used for large area detection. like Figure 4As shown, the specific steps for generating medium-sized and large-sized detection boxes step by step based on the merged group generation algorithm and the initial small-sized target bounding boxes before correction are as follows: Step 2.1: Receive the set of initial small group boxes of the target before correction and the confidence level of each initial small group box of the target before correction; Step 2.2: Select the initial small group box of the target before correction with the highest confidence level as the merging center; Step 2.3: Traverse the remaining initial small group boxes of the target before correction and the intersection-union ratio of the merge center. , will be compared Greater than a given threshold All the initial small group boxes of the target before correction are merged into a single medium group detection box; among them, box A and box B are... Defined as: Step 2.4: If the number of detection boxes in the group reaches... End the merging process and discard the unmerged target initial small group boxes before correction to obtain the final multiple medium group detection boxes. Otherwise, select the target initial small group box before correction with the highest confidence from the unmerged medium group boxes as the merging center and execute step 2.4 again to merge the current unmerged target initial small group boxes before correction. Step 2.5: Sort the medium-group detection boxes according to their confidence scores and take the medium-group detection box with the highest confidence score as the benchmark. Traverse the remaining medium-group detection boxes and merge the medium-group detection boxes whose intersection-union ratio with the benchmark exceeds a given threshold. If the merged region covers the outer range of all medium-group detection boxes, a large-group detection box is obtained. Otherwise, traverse the remaining medium-group detection boxes and continue merging.

[0025] Based on the coordinates of the mid-group and large-group detection boxes, the corresponding regions are cropped from the original SAR remote sensing image. Then, the Gaussian Scaling Adjusting Module (GSAM) is used to process them to ensure that the size of different groups is kept within a reasonable range, that is, neither too large nor too small, thus avoiding the loss of original ship scale information caused by direct cropping or scaling.

[0026] like Figure 5 As shown, the specific implementation steps of the Gaussian size adjustment module are as follows: Calculate the absolute size and average absolute size of all targets within the clipping region; scale the clipping region according to the scaling factor; determine whether padding or clipping is needed after scaling. If the scaled size is too small, perform padding; if the size is too large, perform clipping. Repeat the above operations until the scaled sizes of all target regions are within a reasonable range. Specifically: Step 3.1: Calculate the absolute size of all targets in each clipping region cut out by the middle group detection box and the large group detection box, and calculate the average absolute size based on the absolute size of all targets; The formula for the absolute size of all targets is: in, For the first Within the first cropping area One goal width, For the first Within the first cropping area One goal height, Indicate target The absolute size; The average absolute size is the average of the absolute sizes of all targets within each cut area.

[0027] Step 3.2: Randomly sample the corresponding scaling factor from a Gaussian distribution with the mean absolute size; Step 3.3: Scale the corresponding cropping area according to the scaling factor; Step 3.4: Determine whether each scaled cropped area needs to be filled or cropped. If the scaled cropped area is smaller than the given size, fill it; if the scaled cropped area is larger than the given size, crop it.

[0028] The size-processed medium and large populations are input into the medium-population detector and the large-population detector, respectively. The medium-population detector and the large-population detector share weights to ensure cooperative consistency.

[0029] Step 4: Construct a progressive detector for perception-retrieval-localization, including a guided feature sampling module and a guided bounding box iterative optimization module. The guided feature sampling module is used to sample key features step by step in different groups, and the guided bounding box iterative optimization module is used to iteratively correct the detection box by combining group spatial information. The purpose is to use the guided feature sampling module to sample important features step by step and the guided bounding box iterative optimization module to iteratively correct the detection box step by step, thereby achieving error correction. PRL is a strong backbone detector that uses Deformable-DETR as its model foundation, resulting in higher efficiency and stronger detection performance. Within the group learning framework, PRL consists of two parts: Guided Feature Sampling (PRL-GFS) and Guided Box Iterative Refinement (PRL-GBIR).

[0030] like Figure 6 As shown, the specific implementation steps of the PRL-GFS module, i.e., the guided feature sampling module, are as follows: Step 4.1: Flatten the small group boxes after position correction and the medium and large group boxes after scale correction. Based on the flattening results, generate the query vector and key vector required by the attention mechanism through two linear layers respectively. Step 4.2: The nonlocal neural network obtains the corrected small group boxes, medium group boxes, and large group boxes from the input; Using three 1×1×1 convolution mapping functions , and Generate corresponding query features, key features, and value features respectively. Multiply the query features and key features of each group by the feature similarity weight, and then multiply the feature similarity weight by the value feature to obtain the aggregated features. Using an output mapping The aggregated features are linearly mapped, and the mapped features are added to the synthetic aperture radar remote sensing image to enhance the input synthetic aperture radar remote sensing image. The enhanced image is passed through three residual convolutional layers, then through a linear mapping layer for feature projection, and combined with the key vector to generate the value vector in the attention mechanism; Step 4.3: The query vectors corresponding to each group pass through a linear mapping layer, and then through three attention heads (Head 1, Head 2, and Head 3) corresponding to the corrected large group bounding boxes, medium group bounding boxes, and small group bounding boxes, respectively, to output a set of two-dimensional offsets. ,in, Indicates the first The first attention point in the The first query vector corresponding to the first At each sampling point Direction shift, Indicates the first The first attention point in the The first query vector corresponding to the first At each sampling point The direction is shifted, and at the same time, the query vectors corresponding to each group are normalized into a set of attention weights by the Softmax function after passing through another linear mapping layer. Step 4.4: Construct the feature spaces of the corrected small group boxes, medium group boxes, and large group boxes using the value vectors output in Step 4.2. Then, use three attention heads, Head 1, Head 2, and Head 3, to perform sparse sampling from the feature spaces of the corresponding large group boxes, medium group boxes, and small group boxes to obtain the corresponding salient features. Step 4.5: For each attention head, the corresponding salient features are sampled and aggregated with the learned attention weights to obtain aggregated sampled values. Then, a linear transformation is applied to all aggregated features to output an aggregated feature image.

[0031] The feature images of different group ranges in the aggregated feature map are as follows: in, Represents the query vector. Represents the positional information corresponding to the key vector , This represents the feature map used as input for guided feature sampling, i.e., the corrected small group boxes, medium group boxes, and large group boxes. This indicates the amount of offset, that is... One's attention, This represents the number of sampling points corresponding to each two-dimensional offset. Indicates the first The offset at the ... The first query vector corresponding to the first Attention weights for each sampling point , This represents the learnable weight parameters. This represents the sampling offset constrained within a large population. This represents the sampling offset constrained within the population. This represents the sampling offset constrained within a small group. This represents the feature image of the corrected large bounding box in the output. The output is the feature image of the corrected mid-group bounding box. This represents the feature image of the corrected small group bounding boxes output. As shown in the figure, the specific implementation steps of the PRL-GBIR module, i.e., the guided bounding box iterative optimization module, are as follows: Step 5.1: First, crop the corrected large group bounding box LS, medium group bounding box MS, and small group bounding box SS. Then, perform a flattening operation on the cropped feature map to obtain a feature sequence with uniform feature dimensions. Step 5.2: Align the feature sequences of different groups in dimensions using a fully connected layer; Step 5.3: Concatenate the feature sequences of the aligned large group boxes, medium group boxes, and small group boxes to different layers of the Transformer decoder. Use three iterations to achieve the fusion of group information and decoding process, and finally output a single detection box. The bottom layer decoder incorporates the features corresponding to the large group box, the middle layer decoder incorporates the features corresponding to the medium group box, and the top layer decoder incorporates the features corresponding to the small group box.

[0032] The specific steps for fusing group information and the decoding process using three iterations to finally output a single ship detection box are as follows: First, the aggregated feature image output by the guided feature sampling module is used as the initial predicted bounding box. Under the spatial constraints of the corrected large group bounding boxes (i.e., the constraints imposed by the underlying decoder incorporating the features corresponding to the large group bounding boxes), the first round of bounding box regression is performed to obtain the updated first round of predicted bounding boxes. The formula is as follows: in, This represents the predicted bounding box obtained in the first round. This represents the sigmoid activation function. This represents the inverse sigmoid function. This represents the regression offset in the first round. This refers to the reference box used before the first round of regression begins; Then, using the predicted bounding boxes obtained in the first round as input, a second round of bounding box regression is performed under the corrected mid-group bounding box spatial constraints (i.e., constraints imposed by the mid-layer decoder incorporating the features corresponding to the mid-group bounding boxes), resulting in updated predicted bounding boxes for the second round. The formula is as follows: in, This represents the predicted bounding box obtained in the second round. This represents the regression offset in the second round; Finally, based on the predicted bounding boxes from the second round, a third round of bounding box regression is performed under the spatial constraints of the corrected small group boxes (i.e., the constraints imposed by the top-level decoder incorporating the features corresponding to the small group boxes), resulting in the updated predicted bounding boxes for the third round. The formula is as follows: in, This represents the predicted bounding box obtained in the third round. This represents the regression offset in the third round; Through the above three iterations, the regression offsets of the corrected large group boxes, medium group boxes, and small group boxes are gradually corrected, achieving progressive and accurate localization of the detection boxes, and finally outputting the single ship detection box.

[0033] Step 5: Input the corrected group boxes into the progressive detector to obtain all detection boxes, and sort them in descending order of confidence based on the Box Ensemble Fusion (BEF) algorithm. Select the box with the highest confidence as the master box, traverse the remaining boxes, and compare their intersection-union ratio with the master box. If it is greater than the threshold, perform weighted fusion; otherwise, keep it as a new box. Finally, obtain the predicted group boxes and complete the decision-level fusion. The weights of the weighted fusion are determined by the confidence of each detection box.

[0034] During the training phase, the predicted bounding boxes from steps 2-5 and the ground truth bounding boxes are modeled as two-dimensional Gaussian distributions. The center coordinates of the boundary are used as the mean vector of the Gaussian distribution, and the covariance matrix is ​​constructed using the squares of half the width and height of the bounding box. The training is performed by calculating the second-order Wasserstein distance loss function between the two two-dimensional Gaussian distributions.

[0035] In SAR images, ships are often closely packed or adjacent to each other. To measure the similarity between ground truth and predicted bounding boxes, ship target boxes are modeled as two-dimensional Gaussian distributions, and the differences between boxes are measured by calculating the Wasserstein distance between them. This method better reflects the actual geometry of the ships than directly comparing bounding boxes.

[0036] For a ship, its center is Width is The height is Its two-dimensional Gaussian distribution is defined as follows: , where the mean vector Covariance Matrix The definition is as follows: , The ground reality frame is represented by a two-dimensional Gaussian distribution. The prediction box is represented by a two-dimensional Gaussian distribution as follows: .

[0037] The second-order Wasserstein distance is defined as: in, , These represent the x and y coordinates of the center point of the ground view frame, respectively. , These represent the x and y coordinates of the center point of the prediction box, respectively. , These represent the width and height of the ground view frame, respectively. , These represent the width and height of the prediction box, respectively.

[0038] A new metric is constructed using the second-order Wasserstein distance exponent, which makes training more stable, resulting in the Normalized Wasserstein Distance (NWD): Where C is a hyperparameter used in training. refer to , refer to .

[0039] Experiments were conducted on the SSDD and HRSID datasets, and the specific detection results of the target detection method in this application are shown in Figure 10. This invention achieves the current state-of-the-art accuracy. The AP on the SSDD and HRSID datasets reaches 74.7% and 72.3%, respectively. Incorporating group information at different scales leads to varying degrees of accuracy improvement, fully demonstrating the effectiveness of this invention. As shown in the above data, this invention, based on the concept of group learning, generates an initial small group set through a region candidate network of a weak detector, focusing on the loose spatial correlation information of ship groups, thus avoiding the problem of missed detections of targets in dense scenes by traditional strict detectors. Compared to existing methods that only focus on the independent detection mode of a single target, this application ensures the integrity of the initial detection results by mining the hierarchical spatial information of large, medium, and small groups. Through the PRL-guided feature sampling module and the PRL-guided bounding box iterative optimization module, group spatial constraints are integrated into the detection framework, achieving targeted feature extraction and progressive optimization of the detection box. PRL-GFS improves feature purity by constraining sampling offsets based on group size, filtering out background noise and irrelevant interference. PRL-GBIR iteratively corrects detection box offsets by incorporating group information, simulating the coarse-to-fine cognitive logic of the human visual system to gradually eliminate hierarchical interference. This progressive perception-retrieval-localization model based on group information can handle complex situations in remote sensing images, including scenarios with densely distributed targets, severe background interference, and large differences in target scale. The model's accompanying Gaussian scaling module and box fusion modules exhibit strong synergy and can be further optimized through end-to-end training. Furthermore, the core group learning concept and progressive detection logic can be extended to other remote sensing image target detection tasks.

[0040] Based on the same inventive concept, this application also provides a system (or device) for implementing the aforementioned target detection method for sensing, retrieving, and locating synthetic aperture radar remote sensing images. It can be a server or a terminal, and its internal structure diagram can be as follows. Figure 9As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The system's (or device's) database stores target detection data for sensing, retrieving, and locating targets based on synthetic aperture radar (SAR) remote sensing images. The computer device's I / O interfaces facilitate information exchange between the processor and external devices. The computer device's communication interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a sensing, retrieving, and locating detection method for SAR remote sensing images.

[0041] Those skilled in the art will understand that Figure 10 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0042] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0043] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0044] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0045] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

Claims

1. A method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images, characterized in that, Includes the following steps: Step 1: Acquire the synthetic aperture radar remote sensing image to be detected; Step 2: Input the synthetic aperture radar remote sensing image into the region candidate network of Faster R-CNN to generate the initial small group box of the target, and use a convolutional neural network to correct the position of the initial small group box of the target to obtain the corrected small group box. The convolutional neural network includes a Rol extractor and a classification network connected in sequence. Step 3: Based on the merged group generation algorithm and the initial small group bounding box of the target before correction, generate medium group detection boxes and large group detection boxes step by step. Then, crop out the regions of the synthetic aperture radar remote sensing image corresponding to the medium group detection boxes and large group detection boxes respectively. After the size processing of each cropped region is completed by the Gaussian size adjustment module, the regions are input into the medium group detector and the large group detector respectively for detection to obtain the corrected medium group bounding box and large group bounding box. Step 4: Construct a progressive detector for perception-retrieval-localization, including a guided feature sampling module and a guided bounding box iterative optimization module; Step 5: Input the corrected group boxes into the progressive detector to obtain all detection boxes, and sort them in descending order of confidence based on the box fusion algorithm. Select the box with the highest confidence as the main box, traverse the remaining boxes, and compare the intersection-union ratio with the main box. If it is greater than the threshold, perform weighted fusion; otherwise, keep it as a new box. Finally, the predicted group boxes are obtained.

2. The method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images according to claim 1, characterized in that, During the training phase, the predicted bounding boxes from steps 2-5 and the ground truth bounding boxes are modeled as two-dimensional Gaussian distributions. The center coordinates of the boundary are used as the mean vector of the Gaussian distribution, and the covariance matrix is ​​constructed using the squares of half the width and height of the bounding box. The training is performed by calculating the second-order Wasserstein distance loss function between the two two-dimensional Gaussian distributions.

3. The method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images according to claim 2, characterized in that, In step 3, the specific steps for generating medium-group and large-group detection boxes step by step based on the merged group generation algorithm and the initial small group boxes of the target before correction are as follows: Step 2.1: Receive the set of initial small group boxes of the target before correction and the confidence level of each initial small group box of the target before correction; Step 2.2: Select the initial small group box of the target before correction with the highest confidence level as the merging center; Step 2.3: Traverse the remaining initial small group boxes of the target before correction and the intersection-union ratio of the merge center. , will be compared Greater than a given threshold All the initial small group boxes of the target before correction are merged into a medium group detection box; Step 2.4: If the number of detection boxes in the group reaches... End the merging process and discard the unmerged target initial small group boxes before correction to obtain the final multiple medium group detection boxes. Otherwise, select the target initial small group box before correction with the highest confidence from the unmerged medium group boxes as the merging center and execute step 2.4 again to merge the current unmerged target initial small group boxes before correction. Step 2.5: Sort the medium-group detection boxes according to their confidence scores and take the medium-group detection box with the highest confidence score as the benchmark. Traverse the remaining medium-group detection boxes and merge the medium-group detection boxes whose intersection-union ratio with the benchmark exceeds a given threshold. If the merged region covers the outer range of all medium-group detection boxes, a large-group detection box is obtained. Otherwise, traverse the remaining medium-group detection boxes and continue merging.

4. The target detection method for sensing, retrieving, and locating synthetic aperture radar remote sensing images according to claim 3, characterized in that, The specific implementation steps of the Gaussian size adjustment module in step 3 are as follows: Step 3.1: Calculate the absolute size of all targets in each clipping region cut out by the middle group detection box and the large group detection box, and calculate the average absolute size based on the absolute size of all targets; The formula for the absolute size of all targets is: in, For the first Within the first cropping area One goal width, For the first Within the first cropping area One goal height, Indicate target The absolute size; Step 3.2: Randomly sample the corresponding scaling factor from a Gaussian distribution with the mean absolute size; Step 3.3: Scale the corresponding cropping area according to the scaling factor; Step 3.4: Determine whether each scaled cropped area needs to be filled or cropped. If the scaled cropped area is smaller than the given size, fill it; if the scaled cropped area is larger than the given size, crop it.

5. The method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images according to claim 4, characterized in that, The specific implementation steps of the guided feature sampling module are as follows: Step 4.1: Flatten the small group boxes after position correction and the medium and large group boxes after scale correction. Based on the flattening results, generate the query vector and key vector required by the attention mechanism through two linear layers respectively. Step 4.2: The nonlocal neural network obtains the corrected small group boxes, medium group boxes, and large group boxes from the input; Using three 1×1×1 convolution mapping functions , and Generate corresponding query features, key features, and value features respectively. Multiply the query features and key features of each group by the feature similarity weight, and then multiply the feature similarity weight by the value feature to obtain the aggregated features. Using an output mapping The aggregated features are linearly mapped, and the mapped features are added to the synthetic aperture radar remote sensing image to enhance the input synthetic aperture radar remote sensing image. The enhanced image is passed through three residual convolutional layers, then through a linear mapping layer for feature projection, and combined with the key vector to generate the value vector in the attention mechanism; Step 4.3: The query vectors corresponding to each group pass through a linear mapping layer, and then through three attention heads (Head 1, Head 2, and Head 3) corresponding to the corrected large group bounding boxes, medium group bounding boxes, and small group bounding boxes, respectively, to output a set of two-dimensional offsets. ,in, Indicates the first The first attention point in the The first query vector corresponding to the first At each sampling point Direction shift, Indicates the first The first attention point in the The first query vector corresponding to the first At each sampling point The direction is shifted, and at the same time, the query vectors corresponding to each group are normalized into a set of attention weights by the Softmax function after passing through another linear mapping layer. Step 4.4: Construct the feature spaces of the corrected small group boxes, medium group boxes, and large group boxes using the value vectors output in Step 4.

2. Then, use three attention heads, Head 1, Head 2, and Head 3, to perform sparse sampling from the feature spaces of the corresponding large group boxes, medium group boxes, and small group boxes to obtain the corresponding salient features. Step 4.5: For each attention head, the corresponding salient features are sampled and aggregated with the learned attention weights to obtain aggregated sampled values. Then, a linear transformation is applied to all aggregated features to output an aggregated feature image.

6. The method for sensing, retrieving, and locating targets in synthetic aperture radar remote sensing images according to claim 5, characterized in that, The specific implementation steps of the guided frame iterative optimization module are as follows: Step 5.1: First, crop the corrected large group boxes, medium group boxes, and small group boxes, and then perform a flattening operation on the cropped feature maps to obtain a feature sequence with uniform feature dimensions. Step 5.2: Align the feature sequences of different groups in dimensions using a fully connected layer; Step 5.3: Concatenate the feature sequences of the aligned large group boxes, medium group boxes, and small group boxes to different layers of the Transformer decoder. Use three iterations to achieve the fusion of group information and decoding process, and finally output a single detection box. The bottom layer decoder incorporates the features corresponding to the large group box, the middle layer decoder incorporates the features corresponding to the medium group box, and the top layer decoder incorporates the features corresponding to the small group box.

7. The target detection method for sensing, retrieving, and locating synthetic aperture radar remote sensing images according to claim 3, characterized in that, In step 5.3, the specific steps for fusing group information and the decoding process using three iterations to finally output a single ship detection box are as follows: First, the aggregated feature image output by the guided feature sampling module is used as the initial predicted bounding box. Under the spatial constraints of the corrected large group of bounding boxes, the first round of bounding box regression is performed to obtain the updated first round of predicted bounding boxes, as shown in the formula: in, This represents the predicted bounding box obtained in the first round. This represents the sigmoid activation function. This represents the inverse sigmoid function. This represents the regression offset in the first round, which is the two-dimensional offset of the large group bounding box. This refers to the reference box used before the first round of regression begins, i.e., the aggregated feature image; Then, using the predicted bounding boxes obtained in the first round as input, a second round of bounding box regression is performed under the corrected mid-group bounding box spatial constraints to obtain the updated predicted bounding boxes for the second round. The formula is as follows: in, This represents the predicted bounding box obtained in the second round. This represents the regression offset in the second round, which is the two-dimensional offset of the middle group box; Finally, based on the predicted bounding boxes from the second round, a third round of bounding box regression is performed under the spatial constraints of the corrected small group of bounding boxes to obtain the updated predicted bounding boxes for the third round. The formula is as follows: in, This represents the predicted bounding box obtained in the third round. This represents the regression offset in the third round, which is the two-dimensional offset of the small group bounding box; Through the above three iterations, the regression offsets of the corrected large group boxes, medium group boxes, and small group boxes are gradually corrected, achieving progressive and accurate localization of the detection boxes, and finally outputting the single ship detection box.

8. A target detection system for sensing, retrieving, and locating synthetic aperture radar remote sensing images, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

Images