A system and method for directional target detection for remote sensing images

Abstract

The application relates to the technical field of artificial intelligence, and discloses a directional target detection system and method for remote sensing images. The system comprises: a backbone module configured to extract feature information of a target image and encode to generate first feature data; an aggregation center predictor configured to predict the target number of an aggregation center corresponding to the target image; a decoding module configured to decode the first feature data based on an initial first number of dense queries to obtain a feature set; filter out the target number of aggregation centers from the feature set, and aggregate information of the feature set into each aggregation center; decode based on each aggregation center to obtain an optimized feature set; and an output module configured to generate a prediction result of the target image based on the optimized feature set. The system can improve the adaptability of an end-to-end directional target detection scheme to different remote sensing scenes and improve detection precision.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a directional target detection system and method for remote sensing images. Background Technology

[0002] In recent years, deep learning-based algorithms have been widely applied to various tasks in computer vision, such as image classification, object detection, and object tracking. The end-to-end object detection (DETEction TRansformer, DETR) algorithm introduces the Transformer architecture into the field of object detection. As an important branch of object detection, directional object detection accurately identifies the location, category, and orientation information of targets with arbitrary orientations and significant aspect ratio variations in complex scene images or remote sensing scenes. It has significant application value in remote sensing image interpretation, aerospace, military reconnaissance, port monitoring, and intelligent transportation.

[0003] In high-resolution scenarios such as remote sensing, the number and density of target objects often vary significantly. For example, remote sensing images may contain a large number of small, densely distributed objects, or a small number of sparsely distributed large objects. Current end-to-end directional target detection schemes based on the DETR framework cannot adapt well to remote sensing scenarios with large variations in target density distribution and scale, often resulting in missed detections and false detections. The models exhibit poor generalization ability and low accuracy of detection results, which are problems that need to be addressed. Summary of the Invention

[0004] The purpose of this application is to provide a directional target detection system and method for remote sensing images, so as to improve the adaptability of the end-to-end directional target detection scheme to different remote sensing scenarios, and improve the generalization of the model and the accuracy of the detection results.

[0005] To achieve the above objectives, the technical solution of this application is as follows: In a first aspect, embodiments of this application provide a directional target detection system for remote sensing images, the system comprising: The backbone module is configured to extract feature information from the target image and encode it to generate the first feature data; An aggregation center predictor is configured to predict the number of targets corresponding to aggregation centers in the target image based on the first feature data. The decoding module is configured to decode the first feature data based on an initial first number of dense queries to obtain a feature set containing the first number of dense query features; the first number is greater than the target number; the target number of dense query features are selected from the feature set and determined as aggregation centers; the dense query features in the feature set other than the aggregation centers are merged with each aggregation center, and the merged aggregation centers are decoded to obtain an optimized feature set; The output module is configured to generate a prediction result of the target image based on the optimized feature set; the prediction result includes: the category of the object in the target image and information on the rotated bounding box.

[0006] Optionally, the aggregation center predictor is configured to predict the number of aggregation centers corresponding to the target image based on the first feature data, specifically including: The first feature data is reorganized to obtain a multi-scale feature map; The feature map with the highest resolution is selected from the multi-scale feature maps and determined as the target feature map. The target feature map is classified and predicted by the first prediction head to obtain the range of the number of objects to be detected in the target image. Based on the range of the number of objects to be detected, the target number of aggregation centers is determined.

[0007] Optionally, the decoding module includes: multiple decoding layers, and an aggregation center selector and a query aggregator embedded between adjacent first and second decoding layers; The aggregation center selector is configured to acquire the feature set output by the first decoding layer and the corresponding intermediate prediction results; based on the class confidence and the position of the rotated bounding box, filter out the intermediate prediction results of the target number, and determine the dense query features corresponding to the filtered intermediate prediction results as the aggregation center; The query aggregator is configured to input the merged aggregation center into the second decoding layer.

[0008] Optionally, the aggregation center selector is configured to filter intermediate prediction results for the number of targets based on class confidence and the position of the rotated bounding box, specifically including: Obtain the class confidence score of each intermediate prediction result, and sort the intermediate prediction results from high to low according to the class confidence score; Based on the sorting result, each intermediate prediction result is traversed, the overlap between the rotated bounding box of the currently traversed intermediate prediction result and the rotated bounding box of the previously traversed intermediate prediction results is calculated, and the overlap is compared with a first threshold. If the overlap is less than the first threshold, the intermediate prediction results of the current traversal are filtered out. Stop iterating once intermediate predictions for the target number have been filtered out.

[0009] Optionally, the aggregation center selector, configured to filter intermediate predictions of the number of targets based on class confidence and the position of the rotated bounding box, further includes: If all intermediate prediction results in the sorting results have been traversed and the number of intermediate prediction results selected is less than the target number, the intermediate prediction results with the highest category confidence are selected from the remaining intermediate prediction results in the sorting results to fill in the gaps until the number of intermediate prediction results selected reaches the target number.

[0010] Optionally, the query aggregator is configured to fuse dense query features (excluding aggregation centers) in the feature set with each aggregation center through cross-attention computation, specifically including: Self-attention calculation is performed on all aggregation centers to obtain the first intermediate result; Cross-attention calculation is performed on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain the second intermediate result; The second intermediate result is processed by a feedforward network to obtain the fused aggregation center.

[0011] Optionally, the query aggregator is configured to perform cross-attention calculation on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain a second intermediate result, including: Based on the first intermediate result, the target number of aggregation centers are used as queries in the multi-head attention mechanism, and other dense queries in the feature set other than the aggregation centers are used as keys and values. Cross-attention calculation is performed based on the query matrix and the dense matrix composed of keys and values ​​to obtain the calculation result. The calculation results are subjected to layer normalization to obtain a second intermediate result.

[0012] Optionally, the backbone module includes: The feature extraction module is configured to extract multi-scale feature maps from the target image and generate a feature sequence based on the multi-scale feature maps; The encoding module is configured to perform positional encoding on the feature sequence to generate first feature data.

[0013] Optionally, the system further includes: The visualization module is configured to display the rotated bounding boxes and corresponding category information of each object in the target image based on the prediction results of each object in the target image.

[0014] Secondly, embodiments of this application provide a method for directional target detection based on remote sensing images, applied to the system provided in the first aspect of this application. The method includes: Extract the feature information of the target image and encode it to generate the first feature data; Based on the first feature data, predict the number of targets corresponding to the aggregation center of the target image; The first feature data is decoded based on an initial first number of dense queries to obtain a feature set containing the first number of dense query features; the first number is greater than the target number. The target number of dense query features are selected from the feature set and determined as aggregation centers; The dense query features in the feature set, excluding the aggregation centers, are fused with each aggregation center, and the fused aggregation centers are decoded to obtain the optimized feature set. Based on the optimized feature set, a prediction result for the target image is generated; the prediction result includes: the category of the object in the target image and information on the rotated bounding box.

[0015] The directional target detection system for remote sensing images provided in this application first extracts feature information from the remote sensing image to be processed (i.e., the target image) through a backbone module and encodes it to obtain first feature data. Then, an aggregation center predictor predicts the number of targets that need to be aggregated into the target image based on the first feature data. Further, in the decoding module, based on the predicted number of targets, dense queries for the target number are selected from the dense query features of the first number as aggregation centers, and the information of other dense query features is aggregated into each aggregation center, so that the subsequent decoding process is performed on each aggregated aggregation center, making the number of queries match the distribution density of objects in the remote sensing image. Finally, the output module performs directional target prediction based on the optimized features of the target number to obtain the category and rotated bounding box information of each object in the target image.

[0016] Compared to the traditional DETR framework scheme that uses a fixed number of queries for detection, this application pre-predicts the number of target clusters that match the target image based on the characteristics of the target image using an aggregation center predictor before the decoding stage. Then, in the decoding stage, based on the prediction results of the aggregation center predictor, the initial large number of queries (i.e., the first number of dense queries) are adaptively filtered and feature information is aggregated, so that the number of queries used by the model in the decoding stage and one-to-one matching is adapted to the dynamic changes in the number, distribution density and scale of objects in different remote sensing images.

[0017] Building upon this foundation, for remote sensing images with sparsely distributed objects, this scheme significantly reduces the number of queries, lowers unnecessary computational overhead and redundant prediction results, and improves detection accuracy. Furthermore, by reducing the number of redundant queries, it also reduces the proportion of high-quality negative samples in the one-to-one matching end-to-end training process of the DETR framework, enhancing the utilization of supervision signals, accelerating model convergence, and improving model stability. For remote sensing images with densely distributed objects, it retains a sufficient number of queries for detection, avoiding missed detections and improving detection accuracy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a directional target detection system for remote sensing images proposed in an embodiment of this application; Figure 2 This is a flowchart of a target image prediction process in one embodiment of this application; Figure 3 This is a flowchart of an embodiment of the present application showing how an aggregation center selector filters aggregation centers; Figure 4 This is a flowchart of a query aggregator fusing information between the aggregation center and the remaining dense query features in one embodiment of this application; Figure 5 This is a flowchart of a directional target detection method for remote sensing images proposed in an embodiment of this application. Detailed Implementation

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

[0021] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0022] In the various embodiments of this application, it should be understood that the sequence number of each process described below does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects as detailed in this application.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0025] Current DETR (Deterministic Object Detection) schemes, building upon horizontal object detection, introduce angle parameters into bounding box regression to jointly model the location, scale, and orientation information of objects in an image. The DETR model uses a fixed number of object queries for object detection. However, in remote sensing scenarios with drastic variations in the number, distribution, and scale of objects, this fixed query approach suffers from serious problems of missed detections and false positives. In scenarios with few and sparsely distributed objects, severe overlap between detection boxes easily occurs, resulting in a large number of high-quality negative samples. This significantly increases the difficulty of distinguishing between positive and negative samples during model training, hindering rapid convergence and impacting model robustness. Furthermore, it incurs substantial unnecessary and redundant computational overhead, wasting computational resources and affecting prediction efficiency. Conversely, in scenarios with a large number and dense distribution of objects, insufficient query counts can lead to missed detections, affecting the accuracy of the results. Moreover, the arbitrary directional distribution of objects in remote sensing scenarios makes the model's angle regression unstable, often exhibiting problems such as angle periodicity and boundary discontinuities. Combined with complex background interference, high inter-class appearance similarity, and a large proportion of small objects, this results in low detection accuracy, failing to meet requirements.

[0026] To suppress query redundancy, related DETR extension schemes introduce post-processing operations such as NMS (Non-Maximum Suppression) during the inference phase. However, these schemes break the advantages of the DETR model being end-to-end and post-processing-free, and also cause additional computational overhead and increase the instability of the model, resulting in poor optimization performance.

[0027] This application introduces dense query filtering and feature information aggregation based on image content during the decoding stage, so that the number of queries used by the model to detect the target image is adapted to the number, distribution and scale of objects in the target image. While maintaining the accuracy of end-to-end detection, it improves the model's adaptability to different remote sensing scenarios, speeds up the convergence speed and improves the robustness of the model.

[0028] The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] Figure 1 This is a schematic diagram of a directional target detection system 100 for remote sensing images according to an embodiment of this application. Figure 1 As shown, the system includes: The backbone module 101 is configured to extract feature information from the target image and encode it to generate first feature data; Aggregation center predictor 102 is configured to predict the number of targets corresponding to aggregation centers in the target image based on the first feature data; The decoding module 103 is configured to decode the first feature data based on an initial first number of dense queries to obtain a feature set containing the first number of dense query features; the first number is greater than the target number; the target number of dense query features are selected from the feature set and determined as aggregation centers; the dense query features in the feature set other than the aggregation centers are fused with each aggregation center, and the fused aggregation centers are decoded to obtain an optimized feature set; The output module 104 is configured to generate a prediction result of the target image based on the optimized feature set; the prediction result includes: the category of the object in the target image and information on the rotated bounding box.

[0030] This application addresses the significant imbalance in the number, density, and scale of targets in remote sensing scenarios, providing an end-to-end directional target detection scheme adaptable to different remote sensing scenarios. In this embodiment, feature information from the remote sensing image (i.e., the target image) is first extracted by the backbone module and encoded into one-dimensional feature data (first feature data). Before entering the decoding and one-to-one matching prediction stages, an aggregation center predictor determines the number of targets required for target image detection based on the first feature data, providing prior constraints for subsequent query number compression. Then, after initializing a fixed number (i.e., the first number) of dense queries in the decoding stage, the decoding module filters the initialized dense query features based on the target number output by the aggregation center predictor, obtaining representative dense query features of the target number as aggregation centers. Next, using each aggregation center as a carrier, information from other non-aggregate center dense features semantically related to it is fused, thereby reducing the impact of redundant queries on training and inference while retaining effective fine-grained information, resulting in aggregated target number aggregation centers. Then, the aggregation center of the target quantity is used for subsequent decoding and one-to-one matching prediction operations to finally obtain the directional target prediction result of the target image, outputting the category of each object, as well as the position, size, angle and other information of the rotated bounding box.

[0031] In this embodiment, by introducing learnable object queries during the decoding stage and pre-outputting prior values ​​for the number of queries based on image features before decoding, a small number of representative aggregation centers with more reasonable spatial distribution can be selected from dense queries during subsequent decoding, reducing the number of redundant queries in the model decoding process. Based on this, compression and information aggregation based on the aggregation centers yield queries with a number adapted to the content of the input target image, avoiding problems such as missed detections due to insufficient query counts or false detections due to excessive query counts in traditional DETR schemes. Compared to traditional schemes, this application reduces unnecessary computational overhead, saves computational resources, and improves detection efficiency. Furthermore, since this application's scheme directly embeds query count compression and information aggregation processing into the decoding stage, post-processing is unnecessary, improving detection accuracy, model stability, and generalization while maintaining the end-to-end detection paradigm, and saving computational overhead. This approach enables the adaptive allocation of query quantity based on the content of the target image in complex remote sensing scenarios, enhancing the model's generalization performance. While ensuring detection accuracy, it accelerates model convergence speed and improves model stability and robustness, making it suitable for various remote sensing scenarios such as urban management and industrial inspection.

[0032] As one embodiment of this application, the main module includes: The feature extraction module is configured to extract multi-scale feature maps from the target image and generate a feature sequence based on the multi-scale feature maps; The encoding module is configured to perform positional encoding on the feature sequence to generate first feature data.

[0033] The core modules of the model include a feature extraction module and an encoding module, used to extract features from the target image and process them to obtain one-dimensional feature data. In this embodiment, the feature extraction module includes a CNN network (e.g., a ResNet residual network) to extract feature maps at multiple scales from the target image. After extracting the feature maps, the feature extraction module flattens them into a one-dimensional feature sequence. The encoding module performs positional encoding on this one-dimensional feature sequence to obtain the first feature data used as input to the decoding module.

[0034] As one embodiment of this application, the aggregation center predictor is configured to predict the number of aggregation centers corresponding to the target image based on the first feature data, specifically including: The first feature data is reorganized to obtain a multi-scale feature map; The feature map with the highest resolution is selected from the multi-scale feature maps and determined as the target feature map. The target feature map is classified and predicted by the first prediction head to obtain the range of the number of objects to be detected in the target image. Based on the range of the number of objects to be detected, the target number of aggregation centers is determined.

[0035] Traditional DETR-based detection methods typically employ a fixed number of object queries for decoding and one-to-one matching. However, this strategy suffers from significant errors when the number of objects varies considerably. In remote sensing images, the number of objects can range from extremely few to extremely many, with significant differences in distribution density and scale, making a fixed query approach clearly unsuitable. In sparse scenes, an excessive number of queries generates a large number of candidate predictions that highly overlap with the true targets. Under one-to-one matching, these are forced to participate in training as background samples, forming high-quality negative samples and interfering with gradients. Conversely, in dense scenes, an insufficient number of queries fails to cover all objects, leading to a decrease in model recall.

[0036] In one embodiment, the aggregation center predictor processes the first feature data using a classification prediction approach to predict the number of aggregation centers required for directional target detection in the target image. Specifically, the aggregation center predictor first reorganizes the first feature data to restore it to a multi-scale feature map. Then, it predicts the number of representative aggregation centers to be retained on the feature map, providing a quantitative prior for subsequent query filtering and aggregation, allowing the effective query size participating in matching to adaptively change with the number of targets in the image. To retain as much image detail as possible (including fine-grained information about small objects and dense regions) to improve detection accuracy, after obtaining the multi-scale feature maps, the feature map with the highest resolution (i.e., the largest size) is selected as the target feature map. The target feature map is then processed by a lightweight prediction head to output the predicted number of aggregation centers.

[0037] Considering the wide range and long-tail characteristics of object distribution in remote sensing scenes, directly regressing the precise number of objects at the aggregation center can easily lead to model instability. Therefore, this embodiment adopts a discrete binning classification prediction method. Specifically, the number of objects in the remote sensing image is pre-divided into multiple quantity intervals. The prediction head outputs the classification result of the corresponding quantity interval based on the target feature map, and the model is trained under supervision using cross-entropy loss. After training is complete, the prediction head of the aggregation center predictor outputs the corresponding quantity interval based on the target prediction map, and then determines the precise prior value of the target quantity based on the quantity interval.

[0038] For example, the pre-defined intervals for aggregation centers include: [1, 50], [51, 200], and [201, 400], with corresponding precise values ​​for the target quantity set for each interval: 40, 120, and 300, respectively. When the quantity interval output by the first prediction head is [51, 200], the aggregation center predictor determines the corresponding target quantity to be 120. This target quantity is then used as the quantity constraint for aggregation centers in the decoding stage, thereby achieving resource allocation and image content adaptation in different remote sensing scenarios, improving detection accuracy and resource utilization efficiency, and enhancing the model's generalization ability and stability.

[0039] As one embodiment of this application, the decoding module includes: multiple decoding layers, and an aggregation center selector and a query aggregator embedded between adjacent first and second decoding layers; The aggregation center selector is configured to acquire the feature set output by the first decoding layer and the corresponding intermediate prediction results; based on the class confidence and the position of the rotated bounding box, filter out the intermediate prediction results of the target number, and determine the dense query features corresponding to the filtered intermediate prediction results as the aggregation center; The query aggregator is configured to input the merged aggregation center into the second decoding layer.

[0040] In one embodiment, the decoding module includes multiple decoding layers, an aggregation center selector, and a query aggregator. The aggregation center selector and query aggregator are embedded between two adjacent first and second decoding layers. The decoding module processes the input first feature data and the initialized dense query through a portion of the decoding layers, and the first decoding layer outputs a feature set consisting of a first number of dense query features, as well as intermediate prediction results corresponding to each dense query feature in the feature set. Optionally, in practical applications, if the decoding module contains six decoding layers, the aggregation center selector and query aggregator can be embedded between the third and fourth decoding layers.

[0041] The aggregation center selector selects intermediate predictions based on their class confidence and the position of the rotated bounding boxes. These intermediate predictions have high class confidence for the target number of items and the rotated bounding boxes do not overlap. The dense query features corresponding to these selected intermediate predictions are then designated as aggregation centers, serving as the benchmark for feature fusion in the query aggregator. The query aggregator fuses the dense query features that are not at aggregation centers into semantically relevant aggregation centers and outputs this to the second decoding layer for further decoding of the features at the aggregation centers.

[0042] Figure 2 This is a flowchart illustrating directional target prediction of a target image in one embodiment of this application. For example... Figure 2As shown, the target image is first input into the backbone module. The CNN network extracts features from the target image to generate a multi-scale feature map, which is then flattened and input into the position encoder to generate the first feature data. This first feature data is then fed into the aggregation center predictor and the decoding module. The aggregation center predictor predicts the prior value (i.e., the number of targets) of the aggregation centers needed for the decoding stage based on this first feature data. In the decoding module, a large number of dense queries are initialized. The first feature data and dense queries are processed through the first M decoding layers to enhance the coverage of dense small targets in the target image. The last decoding layer of the first M layers (i.e., the first decoding layer) outputs the feature set and the corresponding intermediate prediction results. The aggregation center selector selects the dense query features representing the number of targets from the feature set based on the intermediate prediction results, using these as aggregation centers. The query aggregator, based on each aggregation center, fuses the information of the remaining dense query features in the feature set into each aggregation center to obtain the fused aggregation centers representing the number of targets. Then, the query aggregator inputs the fused aggregation centers into the second decoding layer (i.e., the first of the remaining N decoding layers) to complete the feature optimization processing of these fused aggregation centers. Finally, the decoding module outputs the optimized feature set (containing the optimized aggregation centers of the target number) to the prediction head (output module) for processing to obtain the prediction result of the target image. The prediction result includes the category of each object in the image, the center position of the rotation box, size and angle information, etc.

[0043] As one embodiment of this application, the aggregation center selector is configured to filter intermediate prediction results of the target quantity based on class confidence and the position of the rotated bounding box, specifically including: Obtain the class confidence score of each intermediate prediction result, and sort the intermediate prediction results from high to low according to the class confidence score; Based on the sorting result, each intermediate prediction result is traversed, the overlap between the rotated bounding box of the currently traversed intermediate prediction result and the rotated bounding box of the previously traversed intermediate prediction results is calculated, and the overlap is compared with a first threshold. If the overlap is less than the first threshold, the intermediate prediction results of the current traversal are filtered out. Stop iterating once intermediate predictions for the target number have been filtered out.

[0044] In one embodiment, the aggregation center selector sorts the category confidence scores of each intermediate prediction result from high to low, obtaining a sorted result. Then, based on the overlap between the rotated boxes corresponding to each intermediate prediction result, the target number of intermediate prediction results are selected from the sorted result. Filtering based on category confidence scores can obtain intermediate prediction results with higher semantic quality, improving detection accuracy. Furthermore, to avoid multiple queries falling into the same region due to filtering solely based on confidence scores, this embodiment also incorporates a class-independent rotated box suppression strategy. This strategy performs filtering based on the overlap between candidate rotated boxes, suppressing nearly duplicate queries with high spatial overlap, thereby improving the spatial diversity and representativeness of the aggregation center.

[0045] Specifically, each intermediate prediction result in the sorted results is traversed, and the overlap between the rotated bounding box of the currently traversed intermediate prediction result and the rotated bounding boxes of the previously traversed intermediate prediction results is calculated. This overlap is then compared to a first threshold. If the overlap is less than the first threshold, the currently traversed intermediate prediction result is selected; if the overlap is greater than or equal to the first threshold, the current intermediate prediction result is skipped, and the traversal continues to the next intermediate prediction result. This operation effectively avoids the generation of high-quality negative samples. Once the target number of intermediate prediction results has been selected, the traversal stops, and the dense lookup features corresponding to the selected target number of intermediate prediction results are used as aggregation centers.

[0046] In this embodiment, by comprehensively analyzing the semantic confidence and spatial overlap of the intermediate prediction results of each dense query in the mid-to-late stages of the decoding phase, class confidence filtering is combined with class-independent rotation suppression filtering. This ensures class detection accuracy while effectively suppressing redundant prediction boxes for the same object, avoiding the generation of high-quality negative samples. This reduces training gradient interference and optimization inefficiency caused by high-quality negative samples in a one-to-one matching mechanism, improves the utilization efficiency of supervision signals, accelerates model convergence, enhances model robustness and detection accuracy, and speeds up detection efficiency. Compared to traditional schemes that rely on post-processing (e.g., NMS), this scheme supports end-to-end processing, adjusting the number of queries within the decoding phase without post-processing.

[0047] As one embodiment of this application, the aggregation center selector is configured to filter intermediate prediction results of the target quantity based on class confidence and the position of the rotated bounding box, and further includes: If all intermediate prediction results in the sorting results have been traversed and the number of intermediate prediction results selected is less than the target number, the intermediate prediction results with the highest category confidence are selected from the remaining intermediate prediction results in the sorting results to fill in the gaps until the number of intermediate prediction results selected reaches the target number.

[0048] In the above embodiment, if the number of intermediate prediction results selected after traversing all intermediate prediction results is insufficient to reach the target number, the remaining intermediate prediction results that were not selected in the sorting results are traversed again, and the intermediate prediction results with the highest confidence are selected to supplement the target number of intermediate prediction results until the target number of intermediate prediction results are selected. Then, the aggregation center selector determines the dense query features corresponding to the selected target number of intermediate prediction results as aggregation centers and outputs them to the query aggregator for feature information fusion processing. By selecting more representative aggregation centers for the target number, redundant computational overhead during model inference is reduced, and duplicate predictions caused by overlapping rotating boxes during training are avoided, accelerating model convergence and improving model stability and robustness.

[0049] Figure 3 This is a flowchart illustrating the aggregation center selector filtering aggregation centers in one embodiment of this application. For example... Figure 3 As shown, in one embodiment, the intermediate prediction results are first sorted from high to low according to their prediction categories to obtain a sorting result. Based on the sorting result, the system iterates through the intermediate prediction results, selecting those with an overlap less than a first threshold based on the overlap between the later-ranked and earlier-ranked predicted rotation boxes. The iteration stops when the number of selected intermediate prediction results meets the target number. If, after iterating through the sorting results, the number of selected intermediate prediction results is less than the target number, then, based on the confidence level from high to low, a corresponding number of intermediate prediction results are selected from the unselected intermediate prediction results in the sorting result to supplement the target number of intermediate prediction results. The dense query features corresponding to each intermediate prediction result are then identified as aggregation centers.

[0050] In one embodiment of this application, the query aggregator is configured to fuse dense query features (excluding aggregation centers) in the feature set with each aggregation center through cross-attention calculation, specifically including: Self-attention calculation is performed on all aggregation centers to obtain the first intermediate result; Cross-attention calculation is performed on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain the second intermediate result; The second intermediate result is processed by a feedforward network to obtain the fused aggregation center.

[0051] In one embodiment, after selecting multiple aggregation centers from the feature set, the remaining dense queries still retain rich fine-grained information, such as small-scale structures, directional changes, and local contextual relationships. To preserve this fine-grained information as much as possible during subsequent decoding, after selecting a target number of aggregation centers, the query aggregator utilizes the semantic and spatial representativeness of the aggregation centers to integrate complementary details from the remaining dense queries. It aggregates, compresses, and fuses semantically and spatially similar query features, aggregating the information of all dense query features not originating from aggregation centers onto the corresponding aggregation centers. This enhances the overall discriminative power of the feature representation and alleviates the training instability caused by high-quality negative samples.

[0052] Figure 4 This is a flowchart illustrating how a query aggregator fuses information between the aggregation center and the remaining dense query features in one embodiment of this application. For example... Figure 4 As shown, in this embodiment, the feature information aggregation process includes three stages: self-attention modeling, cross-attention aggregation, and feedforward enhancement and normalization. The specific feature information fusion process is as follows: (1) Enhance the structure perception ability within the aggregation center through self-attention modeling. Perform self-attention calculation on all aggregation centers to obtain the first intermediate result. : ; in, Presentation layer normalization operation; Indicates that the query is applied to the aggregation center. Multi-head self-attention mechanism; (2) Take the first intermediate result as Q (i.e., Query) and the dense query features of non-aggregation centers as K and V (i.e., Key / Value). Perform cross-attention calculation on the first intermediate result and the dense query features of non-aggregation centers to obtain the second intermediate result. : ; in, Presentation layer normalization operation; This is a feature for dense queries; This is a multi-head cross-attention mechanism between the aggregation center and dense query features; (3) Based on the second intermediate result, the representation ability of the aggregation center is further enhanced by a feedforward network (FFN) to obtain the fused aggregation center. : ; in, Presentation layer normalization operation; This is the second intermediate result; FFN is a feedforward network.

[0053] In this embodiment, the system aggregates query feature information through a query aggregator, converging all feature information around multiple more reliable and representative aggregation centers. This reduces the impact of redundant and repetitive predictions on training and inference from the source. By fusing information from dense query features that were not selected into various aggregation centers, the aggregated features maximize the preservation of detailed information of the target image. This ensures that the model's localization accuracy and orientation sensitivity are not weakened during training due to the aggregation operation, thereby achieving the goal of retaining fine-grained structure and orientation cues while compressing redundant overhead. Therefore, this scheme can ensure stable orientation and localization accuracy even in complex remote sensing scenarios with complex backgrounds, large scale variations, and dense arrays.

[0054] As one embodiment of this application, the query aggregator is configured to perform cross-attention calculation on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain a second intermediate result, including: Based on the first intermediate result, the target number of aggregation centers are used as queries in the multi-head attention mechanism, and other dense queries in the feature set other than the aggregation centers are used as keys and values. Cross-attention calculation is performed based on the query matrix and the dense matrix composed of keys and values ​​to obtain the calculation result. The calculation results are subjected to layer normalization to obtain a second intermediate result.

[0055] In one embodiment, a query aggregator is used to fuse information between dense query features of aggregation centers and non-aggregation centers. The query aggregator uses the aggregation center as the query term and the dense query features of non-aggregation centers as the aggregated information sources. It calculates the relevance weights between the aggregation center and each information source feature using a multi-head attention mechanism. Based on these relevance weights, the dense query features are weighted and fused to obtain an updated aggregation center feature representation. The multi-head attention mechanism calculates the relevance weights based on a scaled dot product and achieves adaptive selection and fusion of dense query features without introducing a preset threshold. Through this flexible and threshold-free fusion method, the aggregation center can focus on absorbing semantically relevant contextual information and suppress interference from irrelevant or redundant features, thereby improving the stability of the feature representation. Furthermore, this method supports end-to-end model training.

[0056] Specifically, the query aggregator will select k (i.e., the target number) aggregation centers. As the query (Q) in the multi-head attention mechanism, the remaining dense query features that are not aggregated centers (excluding the part used for contrastive denoising) Then, using these as key (K) and value (V), the calculation result is obtained. as follows: ; in, This is the first intermediate result; This is a query matrix consisting of all aggregation centers; It is a dense query matrix composed of other dense queries in the feature set, excluding the aggregation center; , , It is a learnable linear projection matrix; Let be the dimension of the attention subspace. Query projection, attention computation, and feature update are all integrated into the above expression; The calculation results are subjected to layer normalization, and the second intermediate result is calculated. as follows: ; in, Presentation layer normalization operation.

[0057] Through cross-attention calculation, the aggregation center can proactively aggregate the most relevant contextual information from dense queries. For each aggregation center... The feature information fusion process is as follows: ; in, This represents the feature representation of the k-th aggregation center obtained after aggregation; The aggregation center feature representation after self-attention computation; , , It is a learnable linear projection matrix; This represents the i-th aggregated query feature, and its corresponding value vector is... ; Indicates the current aggregation center From the perspective of [the previous sentence], the j-th aggregation center competes with all dense query features for similarity calculation, thus ensuring the normalization of attention weights in the dense query dimension; denoted as the dimension of the attention subspace; N represents the number of dense query features without aggregation centers.

[0058] In this embodiment, the query aggregator calculates the relevance between the aggregation center and other dense query features through multi-head attention. A cross-attention mechanism characterizes semantic relevance through scaled dot product attention and performs weighted fusion, enabling each center to focus on the most relevant region and selectively enhance its own features. Unlike simple query averaging, this semantic-driven aggregation operation allows the aggregation center to actively extract the most informative region, achieving efficient semantic-spatial information fusion. Unlike traditional hard selection methods based on fixed IoU or score thresholds, the query aggregator achieves flexible and adaptive information aggregation through multi-head attention. With learnable attention weights, the model can adaptively capture semantic similarity and aggregate valuable information with differentiated weights, thus possessing stronger robustness and generalization ability. Multi-head attention further enhances the model's ability to capture multi-scale semantics and diverse spatial structures. Each attention head models a specific semantic or spatial pattern, thereby enhancing the diversity and discriminativeness of the representation. The query aggregator completes the information aggregation of the aggregation center and updates the aggregation center representation in a residual manner, forming a richer and more detailed feature set, which is then transmitted to the subsequent decoding layer for iterative optimization.

[0059] Finally, the prediction head outputs the category score and the rotated bounding box parameters. During the model training phase, this embodiment uses Hungarian matching to achieve one-to-one alignment between predictions and annotations, and calculates the loss to complete end-to-end training. After training, the inference phase uses the same dynamic aggregation process to achieve more compact output results and fewer duplicate bounding boxes, improving detection accuracy, efficiency, and stability in complex remote sensing scenarios.

[0060] As one embodiment of this application, the system further includes: The visualization module is configured to display the rotated bounding boxes and corresponding category information of each object in the target image based on the prediction results of each object in the target image.

[0061] In one embodiment, the system further includes a visualization module, which is used to visualize the category information and rotated bounding box information of each object in the image based on the prediction results of the target image output by the prediction head, and display them in the user interface to achieve end-to-end visualization interaction with the user.

[0062] The following is a real-world application case of this solution in an industrial inspection scenario.

[0063] In remote sensing images related to industrial inspections, the spatial distribution of objects typically exhibits significant imbalances. For example, within the core operational area, production equipment, transport vehicles, or material storage facilities are often densely concentrated, with relatively similar scales and complex orientation variations. Conversely, in non-operational or buffer areas of the same image, only a few or even none of the objects requiring detection exist. Furthermore, due to variations in shooting height, imaging angle, and industrial facility layout, different targets within the same industrial inspection remote sensing image exhibit significant differences in scale, orientation, and morphology, further exacerbating the contrast between dense and sparse areas. To address these object distribution characteristics in industrial inspection remote sensing scenarios, this scheme is used to train an end-to-end detection model adapted to this type of scenario. The process is as follows: (1) Dataset Construction. The publicly available remote sensing dataset DOTA was used as the main dataset. To adapt to the network input size and improve training efficiency, the original large-format remote sensing image was first sliced, and the original image was cropped into several sub-images according to the preset slice size and overlap rate; at the same time, the original labels were synchronously mapped to the coordinate system of the corresponding sub-images, generating coordinates with the sub-images. Figure 1 A corresponding annotation file is generated to form training sample pairs (sub-images, sub-image annotations). For targets falling at the slice boundary, it is determined whether to retain them according to preset rules (e.g., based on the visibility ratio or area percentage threshold of the target within the slice) to reduce the interference of incomplete targets on training.

[0064] (2) Configure training parameters. The dynamic query aggregation framework of this scheme is built in the MMRote codebase using Rotated-DINO as the baseline. Rotated-DINO follows the baseline implementation of RHINO, using a 5-parameter rotated bounding box representation: {x,y,w,h,r}. The GIoU loss and its matching cost in DINO are replaced with Kullback–Leibler Divergence (KLD) to better adapt to the regression and matching process of the rotated box.

[0065] (3) Model Training. The corresponding image data and its ground truth data are read from the training dataset. During training, to improve the robustness of the model, the image data is augmented, including cropping and horizontal flipping. The processed image data is input into the model for inference. The model outputs the detection results of each object in the image, including: the category information (category confidence) of each target and the corresponding rotated bounding box. The rotated bounding box is represented by five parameters: {x,y,w,h,r}, where x and y are the coordinates of the target center point, w and h are the width and height, and r is the rotation angle. The model output is matched with the ground truth annotations, and the parameters are updated via backpropagation after calculating the loss. The main loss consists of three parts: classification loss, L1 loss for bounding box regression, and KLD loss for rotated box regression / matching (KLD is used instead of GIoU related terms in this scheme). The total loss is the weighted sum of the three losses, with weights set as follows: L1=5, KLD=2, class=1. Backpropagation is performed on the weighted total loss to complete one parameter update. In addition, a simple classification loss is added to the prior prediction part of the aggregation center predictor to briefly classify the order of magnitude of the targets in the current image. After training, the final model is obtained and validated on a validation set.

[0066] The trained model is used to perform directional target detection on remote sensing images of industrial inspections. The image sources include: pre-stored industrial inspection image data in the database, and inspection image data captured in real time by an RGB monocular camera. After obtaining the model's prediction results for the input image, the system plots the bounding box information and category information from the prediction results onto the original inspection image to mark the target location in the scene and label the category information.

[0067] In this embodiment, the system provided by this solution was experimentally validated against several existing directional target detection schemes on the publicly available remote sensing target detection dataset DOTA to evaluate the detection performance and stability of this method in complex remote sensing scenarios. For performance evaluation, commonly used target detection evaluation metrics were used to measure the detection results, including detection accuracy, recall, and localization accuracy, to comprehensively reflect the model's detection performance under different target densities and scales. Specific experimental results are shown in Tables 1 and 2 below. The Improved-D query aggregator - DETR(ours) represents the experimental results using this system. In this embodiment, the ResNet-50 model (R-50) and the Swin-T model were used as feature extraction modules in the backbone network.

[0068] Table 1: Comparison of this solution with related solutions on the DOTA-v1.0 dataset

[0069] Table 2: Comparison of this solution with related solutions on the DOTA-v1.5 dataset

[0070] As shown in Tables 1 and 2, the detection results of this system were compared with those of several Transformer-based multi-object tracking schemes on the DOTA-v1.0 and DOTA-v1.5 test sets. It can be seen that this scheme has a greater advantage in detection accuracy compared to other schemes, reducing computational overhead while maintaining the end-to-end paradigm. Experimental results show that this system can effectively alleviate the performance fluctuation problem caused by changes in the number of objects in the target image while maintaining high detection accuracy.

[0071] Based on the same inventive concept, one embodiment of this application provides a method for directional target detection in remote sensing images. Figure 5 This is a flowchart of a directional target detection method for remotely sensed images proposed in an embodiment of this application. Figure 5 As shown, the method includes: S1: Extract the feature information of the target image and encode it to generate the first feature data; S2: Based on the first feature data, predict the number of targets corresponding to the aggregation center of the target image; S3: Decode the first feature data based on the initial first number of dense queries to obtain a feature set containing the first number of dense query features; the first number is greater than the target number; S4: Select the target number of dense query features from the feature set and determine them as aggregation centers; S5: Merge the dense query features in the feature set except for the aggregation center with each aggregation center, and decode the merged aggregation center to obtain the optimized feature set. S6: Based on the optimized feature set, generate a prediction result for the target image; the prediction result includes: the category of the object in the target image and information on the rotated bounding box.

[0072] As one embodiment of this application, predicting the number of targets corresponding to the aggregation centers of the target image based on the first feature data specifically includes: The first feature data is reorganized to obtain a multi-scale feature map; The feature map with the highest resolution is selected from the multi-scale feature maps and determined as the target feature map. The target feature map is classified and predicted by the first prediction head to obtain the range of the number of objects to be detected in the target image. Based on the range of the number of objects to be detected, the target number of aggregation centers is determined.

[0073] As one embodiment of this application, selecting the target number of dense query features from the feature set and determining them as aggregation centers includes: Obtain the intermediate prediction results corresponding to the feature set output by the first decoding layer; Based on the category confidence and the position of the rotated bounding box, intermediate prediction results for the target quantity are selected, and the dense query features corresponding to the selected intermediate prediction results are determined as aggregation centers.

[0074] As one implementation of this application, intermediate prediction results for the number of targets are filtered based on category confidence and the position of the rotated bounding box, specifically including: Obtain the class confidence score of each intermediate prediction result, and sort the intermediate prediction results from high to low according to the class confidence score; Based on the sorting result, each intermediate prediction result is traversed, the overlap between the rotated bounding box of the currently traversed intermediate prediction result and the rotated bounding box of the previously traversed intermediate prediction results is calculated, and the overlap is compared with a first threshold. If the overlap is less than the first threshold, the intermediate prediction results of the current traversal are filtered out. Stop iterating once intermediate predictions for the target number have been filtered out.

[0075] As one embodiment of this application, the intermediate prediction results for the number of targets are filtered based on the category confidence level and the position of the rotated bounding box, and further include: If all intermediate prediction results in the sorting results have been traversed and the number of intermediate prediction results selected is less than the target number, the intermediate prediction results with the highest category confidence are selected from the remaining intermediate prediction results in the sorting results to fill in the gaps until the number of intermediate prediction results selected reaches the target number.

[0076] As one implementation of this application, the dense query features in the feature set, excluding the aggregation centers, are fused with each aggregation center, specifically including: Self-attention calculation is performed on all aggregation centers to obtain the first intermediate result; Cross-attention calculation is performed on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain the second intermediate result; The second intermediate result is processed by a feedforward network to obtain the fused aggregation center.

[0077] As one embodiment of this application, cross-attention calculation is performed on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain a second intermediate result, including: Based on the first intermediate result, the target number of aggregation centers are used as queries in the multi-head attention mechanism, and other dense queries in the feature set other than the aggregation centers are used as keys and values. Cross-attention calculation is performed based on the query matrix and the dense matrix composed of keys and values ​​to obtain the calculation result. The calculation results are subjected to layer normalization to obtain a second intermediate result.

[0078] As one embodiment of this application, feature information of the target image is extracted and encoded to generate first feature data, including: Multi-scale feature maps are extracted from the target image, and a feature sequence is generated based on the multi-scale feature maps. The feature sequence is positionally encoded to generate first feature data.

[0079] As one embodiment of this application, after generating the prediction result of the target image, the method further includes: Based on the prediction results of each object in the target image, the rotated bounding box of each object and its corresponding category information are displayed in the target image.

[0080] The specific manner in which each step of the method in the above embodiments is performed has been described in detail in the embodiments of the relevant system, and will not be elaborated here. The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0081] For the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and components involved are not necessarily essential to this application.

[0082] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] Although preferred embodiments of the embodiments of this application have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, this application is to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.

[0087] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0088] The above provides a detailed description of the directional target detection system and method for remote sensing images provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A directional target detection system for remote sensing images, characterized in that, include: The backbone module is configured to extract feature information from the target image and encode it to generate the first feature data; An aggregation center predictor is configured to predict the number of targets corresponding to aggregation centers in the target image based on the first feature data. The decoding module is configured to decode the first feature data based on an initial first number of dense queries to obtain a feature set containing the first number of dense query features; the first number is greater than the target number; the target number of dense query features are selected from the feature set and determined as aggregation centers; the dense query features in the feature set other than the aggregation centers are merged with each aggregation center, and the merged aggregation centers are decoded to obtain an optimized feature set; The output module is configured to generate a prediction result of the target image based on the optimized feature set; the prediction result includes: the category of the object in the target image and information on the rotated bounding box.

2. The directional target detection system for remote sensing images according to claim 1, characterized in that, The aggregation center predictor is configured to predict the number of aggregation centers corresponding to the target image based on the first feature data, specifically including: The first feature data is reorganized to obtain a multi-scale feature map; The feature map with the highest resolution is selected from the multi-scale feature maps and determined as the target feature map. The target feature map is classified and predicted by the first prediction head to obtain the range of the number of objects to be detected in the target image. Based on the range of the number of objects to be detected, the target number of aggregation centers is determined.

3. The directional target detection system for remote sensing images according to claim 2, characterized in that, The decoding module includes: multiple decoding layers, and an aggregation center selector and a query aggregator embedded between adjacent first and second decoding layers; The aggregation center selector is configured to acquire the feature set output by the first decoding layer and the corresponding intermediate prediction results; based on the class confidence and the position of the rotated bounding box, filter out the intermediate prediction results of the target number, and determine the dense query features corresponding to the filtered intermediate prediction results as the aggregation center; The query aggregator is configured to input the merged aggregation center into the second decoding layer.

4. The directional target detection system for remote sensing images according to claim 3, characterized in that, The aggregation center selector is configured to filter intermediate prediction results for the number of targets based on class confidence and the position of the rotated bounding box, specifically including: Obtain the class confidence score of each intermediate prediction result, and sort the intermediate prediction results from high to low according to the class confidence score; Based on the sorting result, each intermediate prediction result is traversed, the overlap between the rotated bounding box of the currently traversed intermediate prediction result and the rotated bounding box of the previously traversed intermediate prediction results is calculated, and the overlap is compared with a first threshold. If the overlap is less than the first threshold, the intermediate prediction results of the current traversal are filtered out. Stop iterating once intermediate predictions for the target number have been filtered out.

5. The directional target detection system for remote sensing images according to claim 4, characterized in that, The aggregation center selector is configured to filter intermediate prediction results for the number of targets based on class confidence and the position of the rotated bounding box, and further includes: If all intermediate prediction results in the sorting results have been traversed and the number of intermediate prediction results selected is less than the target number, the intermediate prediction results with the highest category confidence are selected from the remaining intermediate prediction results in the sorting results to fill in the gaps until the number of intermediate prediction results selected reaches the target number.

6. The directional target detection system for remote sensing images according to claim 3, characterized in that, The query aggregator is configured to fuse dense query features (excluding aggregation centers) from the feature set with each aggregation center, specifically including: Self-attention calculation is performed on all aggregation centers to obtain the first intermediate result; Cross-attention calculation is performed on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain the second intermediate result; The second intermediate result is processed by a feedforward network to obtain the fused aggregation center.

7. The directional target detection system for remote sensing images according to claim 6, characterized in that, The query aggregator is configured to perform cross-attention calculation on the first intermediate result and the dense query features in the feature set excluding the aggregation center to obtain a second intermediate result, including: Based on the first intermediate result, the target number of aggregation centers are used as queries in the multi-head attention mechanism, and other dense queries in the feature set other than the aggregation centers are used as keys and values. Cross-attention calculation is performed based on the query matrix and the dense matrix composed of keys and values ​​to obtain the calculation result. The calculation results are subjected to layer normalization to obtain a second intermediate result.

8. The directional target detection system for remote sensing images according to claim 1, characterized in that, The backbone module includes: The feature extraction module is configured to extract multi-scale feature maps from the target image and generate a feature sequence based on the multi-scale feature maps; The encoding module is configured to perform positional encoding on the feature sequence to generate first feature data.

9. The directional target detection system for remote sensing images according to claim 1, characterized in that, Also includes: The visualization module is configured to display the rotated bounding boxes and corresponding category information of each object in the target image based on the prediction results of each object in the target image.

10. A method for directional target detection in remote sensing images, characterized in that, Applied to the system as described in any one of claims 1-9, comprising: Extract the feature information of the target image and encode it to generate the first feature data; Based on the first feature data, predict the number of targets corresponding to the aggregation center of the target image; The first feature data is decoded based on an initial first number of dense queries to obtain a feature set containing the first number of dense query features; the first number is greater than the target number. The target number of dense query features are selected from the feature set and determined as aggregation centers; The dense query features in the feature set, excluding the aggregation centers, are fused with each aggregation center, and the fused aggregation centers are decoded to obtain the optimized feature set. Based on the optimized feature set, a prediction result for the target image is generated; the prediction result includes: the category of the object in the target image and information on the rotated bounding box.

Images