A remote sensing image target detection method and device resistant to complex background interference

Abstract

The application discloses a remote sensing image target detection method and device resisting complex background interference. The method comprises the following steps: extracting multi-scale visual features of an original input image; performing context enhancement modeling on the multi-scale visual features, constructing a hierarchical specific global context representation, and obtaining an enhanced visual feature representation containing scene context information through attention interaction; further, marking the enhanced candidate target region representation back to the candidate target region features, adaptively updating the candidate target region features, and thus obtaining a context-aware target representation; finally, passing the enhanced features through a detection prediction head, and using classification and regression branches to obtain target category and rotation bounding box predictions respectively. The device comprises a processor and a memory. The application adopts a fusion method based on multi-scale context enhancement and feature feedback update, which can effectively alleviate the missed detection and false detection problems caused by factors such as similar background texture, significant scale change and dense target distribution in complex remote sensing scenes, thereby improving the classification accuracy and positioning precision of remote sensing image target detection.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image target detection technology, and in particular to a remote sensing image target detection method and apparatus that resists interference from complex backgrounds. Background Technology

[0002] With the rapid development of Earth observation satellite, airborne remote sensing, and UAV imaging technologies, high-resolution remote sensing images have been widely applied in scenarios such as land resource monitoring, urban planning, traffic management, maritime supervision, disaster assessment, and national defense security. In these applications, target detection is one of the fundamental tasks of intelligent interpretation of remote sensing images. Its goal is to automatically locate and identify typical ground features such as aircraft, ships, vehicles, bridges, and port facilities from large-format remote sensing images, providing support for subsequent situational analysis and decision-making. Compared to natural scene images, targets in remote sensing images often exhibit significant scale differences, dense distribution, arbitrary orientation, and large aspect ratio variations. Furthermore, due to factors such as imaging perspective, ground feature structure, and lighting and shadows, targets often appear in a rotated state. To more accurately describe the direction and shape of targets, remote sensing image target detection methods typically use rotated bounding boxes to locate targets, enabling the detection results to better fit the target outline and improve the discrimination ability in dense scenes. Therefore, rotated target detection has become an important research direction in the field of remote sensing target detection.

[0003] Most existing methods for rotating target detection are based on deep learning frameworks. Their basic process typically includes steps such as multi-scale feature extraction, candidate region generation or dense prediction, region feature alignment and enhancement, and classification and regression prediction. Depending on the detection paradigm, these methods can be divided into two categories: two-stage detection methods, which typically generate candidate regions first, and then perform feature alignment and fine-grained classification and regression on these regions; and single-stage detection methods, which typically perform dense classification and rotation box regression directly on multi-scale feature maps. To meet the localization requirements of rotating targets, existing methods have made significant improvements in rotation box parameterization, angle regression strategies, positive and negative sample allocation, and feature alignment methods, and have achieved some progress on public datasets.

[0004] However, in complex remote sensing scenarios, background interference and varying imaging conditions still significantly constrain the target detection performance of remote sensing images. Specifically, the background of remote sensing images typically contains a large number of textured and structurally similar features (such as roads, rooftops, docks, container yards, etc.), accompanied by shadow occlusion, resolution variations, and imaging noise. This easily leads to targets appearing similar to the background or having unclear boundaries, resulting in missed detections and false detections. Meanwhile, remote sensing targets often exhibit large scale spans, dense arrangements, and wide rotation angle distributions. Mutual occlusion or close proximity of targets further weakens the separability and localization stability of candidate regions, making the model more prone to class confusion and rotation bounding box regression bias in complex backgrounds. Existing methods mostly focus on local region appearance modeling, insufficiently utilizing cross-scale contextual information and spatial layout cues, making it difficult to stably aggregate effective discriminative information in strongly disturbed backgrounds.

[0005] Therefore, how to overcome the limitations of relying solely on local appearance features for discrimination under conditions of complex background textures, numerous interfering objects, and variable target rotation postures in remote sensing images, and more fully integrate multi-scale contextual clues to achieve more robust detection and localization of rotating targets against complex background interference has become an important technical challenge that urgently needs to be addressed in the field of remote sensing rotating target detection. Summary of the Invention

[0006] This invention provides a method and apparatus for target detection in remote sensing images that resists interference from complex backgrounds. The feature enhancement and fusion method employed in this invention can effectively alleviate the problems of missed detections and false detections in complex remote sensing scenes, and improve the classification accuracy and positioning precision of target detection in remote sensing images. Details are described below: Extract multi-scale visual features from the original input image; Context-enhanced modeling is performed on multi-scale visual features to construct hierarchical global contextual representations, and enhanced visual feature representations containing contextual information are obtained through attention interactions. The enhanced candidate target region label representation is injected back into the candidate target region feature, and the candidate target region feature is adaptively updated to obtain a context-aware target representation. The enhanced features are passed through the detection prediction head, and the parameters of the target category and the rotated bounding box are obtained by using the classification and regression branches, respectively.

[0007] Specifically, the extraction of multi-scale visual features from the original input image includes: A stripNet network pre-trained on a large image classification dataset is used as a visual encoder to extract multi-scale visual features from the input image.

[0008] Specifically, the context-enhanced modeling of multi-scale visual features involves: For each level of visual feature map, gridded pooling downsampling is performed to form a global context feature representation for the corresponding level. Based on this, the RoI (Region of Interest) features corresponding to the candidate target are obtained according to the feature level to which the candidate target belongs. The RoI features and the global context features of the same level are then input into the attention interaction module for feature interaction fusion to obtain an enhanced visual feature representation containing contextual information. Furthermore, residual fusion and feedforward transformation are performed on the enhanced visual feature representation to improve feature representation capability and training stability.

[0009] The method for constructing the hierarchical-specific global context representation is as follows: In the formula, The first output of the feature pyramid network Each level of feature map; This is a gridded adaptive average pooling operation used to... downsampling Grid features; This is an operation that unfolds according to a spatial grid; For the first A hierarchical global context representation, which includes Each context tag is used to represent the global semantic and spatial layout information of this level.

[0010] Furthermore, the method for obtaining enhanced visual feature representations containing contextual information through attention interaction is as follows: In the formula, RoI features for candidate target regions; RoI feature tokenization operation, used to... Mapped to RoI tag representation ; To correspond with the level to which the RoI belongs The corresponding hierarchical global context representation; For multi-head attention operations; For query, As key, Value; For context enhancement; The enhanced RoI tag representation after fusing contextual information, and the above It serves as an enhanced visual feature representation for subsequent candidate target region feature updates and detection predictions.

[0011] Specifically, the step of injecting the enhanced candidate target region labeling representation back into the candidate target region features involves: The candidate target region label representation, enhanced through attention interaction and global context, is returned to the original target region feature map. An adaptive update mechanism is then used to fuse the enhanced label representation with the original target region features, resulting in a more accurate and robust target representation, further improving the target features' ability to discriminate against complex background interference.

[0012] The method for adaptively updating the features of the candidate target region is as follows: In the formula, An enhanced RoI tag representation after incorporating contextual information; This is a normalization operation; It is a linear mapping used to generate modulation parameters; and These are the scale modulation term and the bias modulation term, respectively. RoI features for candidate target regions; This is element-wise multiplication; To modulate the updated structured enhanced RoI features, and to... This serves as the feature input for candidate target regions in subsequent detection and prediction.

[0013] Specifically, the process of passing the enhanced features through a detection prediction head and using classification and regression branches to obtain predictions of the target category and the rotated bounding box involves: The candidate target region features, after context enhancement and adaptive updating, are input into the detection prediction head. Combined with global context information, a collaborative feature representation for target category discrimination and rotation bounding box regression is obtained. Subsequently, the collaborative feature representation is input into the detection prediction head, wherein the classification branch outputs the target category confidence score, and the regression branch outputs the target's rotation bounding box parameters, including center point coordinates, width, height, and rotation angle.

[0014] A remote sensing image target detection device resistant to complex background interference, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to execute any of the methods described above.

[0015] A computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform any of the methods described above.

[0016] The beneficial effects of the technical solution provided by this invention are: (1) In the feature extraction stage, the present invention introduces a multi-scale context enhancement mechanism to construct a corresponding global context representation for visual features at different levels, and adaptively integrates scene context information into candidate target region features through attention interaction; therefore, the obtained target features not only contain local texture and edge information, but also integrate context clues related to scene structure, which can improve the distinguishability of target representation in the case of complex background texture and similar interference objects, thereby reducing false detection and false detection caused by complex background. (2) In the candidate target region feature enhancement stage, this invention introduces a label representation back-injection and adaptive update mechanism. By back-injecting the enhanced candidate target region label representation into the original candidate target region features, the target features can adaptively integrate more contextual information. In this way, the model can combine multi-level global contextual information to refine the target region features, thereby improving the perception ability of target representation. Especially in scenarios with complex background interference, dense target distribution, and large rotational attitude changes, it can significantly improve the stability and accuracy of target detection and reduce false detections and missed detections caused by cluttered backgrounds or blurred targets. Attached Figure Description

[0017] Figure 1 A flowchart of a remote sensing image target detection method that resists interference from complex backgrounds; Figure 2 This is a flowchart of a remote sensing image target detection algorithm that is resistant to complex background interference. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0019] Example 1 This invention provides a remote sensing image target detection method that resists interference from complex backgrounds. See [link to relevant documentation]. Figure 1 The method includes the following steps: 101: Extracting multi-scale visual features from the original input image; For a given raw remote sensing image data, a stripNet network pre-trained on a large image classification dataset is used as a visual encoder to extract multi-scale visual features of the input image. 102: Perform context-enhanced modeling on multi-scale visual features, construct hierarchical global context representations, and obtain enhanced visual feature representations containing contextual information through attention interactions; For each level of visual feature map, gridded pooling downsampling is performed to form a global context feature representation for the corresponding level. Based on this, the RoI (Region of Interest) features corresponding to the candidate target are obtained according to the feature level to which the candidate target belongs. The RoI features and the global context features of the same level are then input into the attention interaction module for feature interaction fusion to obtain an enhanced visual feature representation containing contextual information. Furthermore, residual fusion and feedforward transformation are performed on the enhanced visual feature representation to improve feature representation capability and training stability.

[0020] 103: The enhanced candidate target region label representation is injected back into the candidate target region features, and the candidate target region features are adaptively updated to obtain a context-aware target representation. The candidate target region label representation, enhanced through attention interaction and global context, is returned to the original target region feature map. An adaptive update mechanism is then used to fuse the enhanced label representation with the original target region features, resulting in a more accurate and robust target representation, further improving the target features' ability to discriminate against complex background interference.

[0021] 104: The enhanced features are passed through the detection prediction head, and the parameters of the target category and the rotated bounding box are obtained by using the classification and regression branches, respectively.

[0022] The candidate target region features, after context enhancement and adaptive updating, are input into the detection prediction head. Combined with global context information, a collaborative feature representation for target category discrimination and rotation bounding box regression is obtained. Subsequently, the collaborative feature representation is input into the detection prediction head, wherein the classification branch outputs the target category confidence score, and the regression branch outputs the target's rotation bounding box parameters, including center point coordinates, width, height, and rotation angle.

[0023] Example 2 The following section provides a further explanation of the scheme in Example 1, using specific examples and calculation formulas. See the description below for details: 201: Extract multi-scale visual features from the original input image using a feature extractor; Given an input remote sensing image ,in To determine the image resolution, a StripNet network pre-trained on ImageNet is used as the visual encoder to perform multi-stage feature extraction on the input image, resulting in a multi-scale feature map set. Furthermore, a feature pyramid network is used to perform top-down fusion and lateral connections on multi-scale features, generating pyramid features with a unified channel dimension. ,in , Indicates a hierarchical index. The pyramid feature represents the number of channels. Used for subsequent candidate rotation region generation and region feature alignment.

[0024] 202: Construct hierarchical-specific global contextual representations for multi-scale visual features and achieve contextual enhancement; The feature maps at each level of the feature pyramid obtained in the preceding steps are subjected to gridded adaptive pooling. This invention posits that shallow features emphasize texture and edge information, while deep features emphasize semantic and structural information; however, interference in complex remote sensing contexts often exists simultaneously at different scales and semantic levels. Therefore, for each level of feature map, a fixed grid size is used for downsampling, compressing the global scope information of that level into several spatially located contextual units, thereby constructing a level-specific global context representation. This allows subsequent candidate target regions to reference scene cues of the same scale and semantic level, avoiding semantic mismatch caused by cross-level mixing.

[0025] Next, the candidate rotation regions are assigned to the most matching feature level based on their scale, and rotation alignment operations are performed on the corresponding feature maps to obtain the RoI features of the candidate targets. Furthermore, to facilitate interaction with the context unit, the RoI features are compressed through convolution, pooling, and linear mapping to obtain the RoI label representation corresponding to the candidate targets.

[0026] Finally, the RoI tag representation and the global context representation at the same level are input into the attention interaction module. This allows the RoI tag to adaptively select the most relevant region information from the global context unit in a "query" manner and complete information aggregation, thereby obtaining an enhanced visual feature representation containing contextual information. To avoid destroying the original discriminative cues during the enhancement process, residual fusion is used to combine the enhancement result with the original RoI tag representation, and feedforward transformation is used to further improve feature representation ability and training stability, outputting candidate target region enhancement features for subsequent structured enhancement and detection prediction.

[0027] 203: The enhanced candidate target region label representation is injected back into the candidate target region features to generate a context-aware target representation; Specifically, based on the obtained context-enhanced candidate target region features, this invention introduces an adaptive feature modulation mechanism to further improve the discrimination stability of candidate targets under complex background interference conditions. The core idea is to inject the enhanced candidate target region label representation back into the original candidate target region features, and then apply adaptive modulation to the candidate target region features, so that the target feature representation, while enhancing the context, possesses more robust discrimination capabilities.

[0028] First, based on the enhanced candidate target region feature representation, it is fused with the original features, and adaptive modulation parameters are generated through feature modulation. Then, an adaptive modulation mechanism is used to dynamically update the candidate target region features, enabling the target features to possess more accurate target perception capabilities after incorporating global context information.

[0029] Specifically, the modulation process is achieved through scaling and bias modulation, and its core update form is as follows: in, Represents the original RoI features of the candidate target region. and These represent the scale modulation term and the bias modulation term generated from the structured target representation, respectively. For element-wise multiplication, The features of the candidate target region after structure enhancement.

[0030] Through the aforementioned feature enhancement modeling process, the fusion of contextual information and candidate target region features not only improves the target features' ability to discriminate against complex background interference but also enhances the adaptability of target region features to changes in target scale and rotational attitude. Ultimately, this adaptively modulated target representation provides a more stable and discriminative feature foundation for subsequent target classification and rotational bounding box regression.

[0031] 204: Input the enhanced features into the detection prediction head, and output the target category and the rotated bounding box; This embodiment employs a two-stage rotating target detection framework for final target prediction. First, the enhanced features are input into the rotating target detection prediction head. The classification branch classifies candidate targets and outputs the target category confidence score for each candidate region. Simultaneously, the regression branch regresses the rotation bounding box parameters of the candidate target regions to obtain the target's rotation localization result.

[0032] Specifically, the detection prediction head includes a classification branch and a regression branch. The classification branch predicts the probability distribution of candidate target regions belonging to each target category; the regression branch predicts the rotation bounding box parameters of the candidate targets, which include at least the target center point coordinates, width, height, and rotation angle. The detection prediction process can be represented as follows: in, To enhance the features of candidate target regions in a structured way, For classification prediction head, For the rotation bounding box regression prediction head, For the target category prediction results, This is the prediction result for the rotated bounding box.

[0033] Furthermore, to eliminate redundant detection results among multiple candidate targets, non-maximum suppression processing based on rotation intersection-union ratio is performed on the predicted rotated bounding box and its corresponding category confidence, retaining detection results with high confidence and low spatial overlap, to obtain the final rotated target detection output.

[0034] During the training phase, the optimization objective of the detection prediction head is composed of both the classification loss and the rotation bounding box regression loss, and its overall loss function is: in, Classification loss for the target, For the regression loss of the rotated bounding box, and These are the loss weight coefficients, used in the optimization process to balance classification and localization tasks.

[0035] Example 3 The effectiveness of the method of the present invention will be evaluated below with specific experimental data, as detailed in the following description: 301: Data Composition The dataset used for evaluation is the DOTA remote sensing rotating target detection dataset. The DOTA dataset contains 2,806 high-resolution remote sensing images, covering 15 typical ground target classes, and uses bounding boxes rotated in any direction for accurate annotation. According to the official breakdown, the dataset includes 1,403 training images, 468 validation images, and 935 test images. The training and validation sets are used for model training and parameter tuning, while the test set is used for performance evaluation.

[0036] The experiments were conducted using both single-scale and multi-scale datasets, both derived from the DOTA dataset. In the single-scale setting, all images were trained and tested at a uniform scale. In the multi-scale setting, multi-scale training samples were constructed by applying different scaling ratios to the input images to evaluate the model's performance in rotating object detection scenarios with complex scale variations.

[0037] 302: Evaluation Criteria In both single-scale and multi-scale dataset settings, the mean accuracy (mAP@0.5) at an IoU threshold of 0.5 in the rotated target detection task was used as the main evaluation metric to measure the overall detection performance of the model.

[0038] The Intersection over Union (IoU) ratio between rotated bounding boxes is defined as the ratio of the area of ​​the overlap between the predicted rotated bounding box and the actual rotated bounding box to the area of ​​their union. In the formula, This represents the predicted rotating bounding box. This represents a true rotating bounding box.

[0039] 303: Comparison Algorithm To verify the effectiveness and versatility of the proposed remote sensing image target detection method against complex background interference, comparative experiments were conducted with several mainstream rotating target detection algorithms, including RTMDet-R, LSKNet-S, KLD, and R3Det-KLD, using single-scale and multi-scale settings on the DOTA dataset. All comparison methods were trained and tested under the same data partitioning and evaluation criteria (mAP@0.5) to ensure the fairness of the experimental results. The training and validation sets were used together for model training, while the test set was used only for final performance evaluation.

[0040] Table 1 presents a performance comparison of the proposed method with several existing rotating target detection methods on the DOTA dataset under single-scale and multi-scale settings. As shown in Table 1, significant differences in detection performance exist among the compared methods under complex remote sensing backgrounds and substantial target scale variations. Traditional rotating target detection methods, represented by KLD and R3Det-KLD, exhibit relatively low overall performance under both single-scale and multi-scale settings, indicating that relying solely on local features or angle modeling strategies is insufficient to address the challenges posed by complex background interference and scale variations. In contrast, methods such as RTMDet-R and LSKNet-S achieve some performance improvement under multi-scale settings by introducing stronger backbone networks and feature modeling mechanisms, but still face certain performance bottlenecks under complex backgrounds. After introducing the proposed method for rotating target detection against complex background interference, the mAP@0.5 is improved to 78.87% under single-scale settings and further improved to 82.10% under multi-scale settings, both outperforming existing comparative methods and demonstrating the stable detection capability of the proposed method under different scale conditions.

[0041] Especially in multi-scale settings, the proposed method achieves performance improvements of 0.77% and 0.46% compared to RTMDet-R and LSKNet-S, respectively. This demonstrates that the proposed context enhancement and structured target representation mechanism can effectively suppress background interference and enhance the discriminative power of target region features in remote sensing scenes with significant scale variations and complex background textures, thereby significantly improving the localization accuracy and detection reliability of rotating targets. Furthermore, in single-scale settings, the proposed method maintains high detection accuracy without introducing additional performance degradation, indicating good stability and generalization ability.

[0042] Table 1 Therefore, through experimental verification using a database, the algorithm proposed in this invention can effectively enhance the context and structure the representation of candidate target region features under conditions of complex remote sensing background texture interference and significant scale changes, thereby making the target features more discriminative and stable, thus reducing the risk of false detection and false negative detection caused by similar background textures, demonstrating good versatility and engineering application value.

[0043] Example 4 A remote sensing image target detection device resistant to complex background interference, the device includes: a processor and a memory, the memory storing program instructions, and the processor calling the program instructions stored in the memory to cause the device to execute the following method steps in Embodiment 1: This invention provides a method and apparatus for target detection in remote sensing images that resists interference from complex backgrounds. The feature enhancement and fusion method employed in this invention can effectively alleviate the problems of missed detections and false detections in complex remote sensing scenes, and improve the classification accuracy and positioning precision of rotating target detection. See the description below for details: Extract multi-scale visual features from the original input image; Context-enhanced modeling is performed on multi-scale visual features to construct hierarchical global contextual representations, and enhanced visual feature representations containing contextual information are obtained through attention interactions. The enhanced candidate target region label representation is injected back into the candidate target region feature, and the candidate target region feature is adaptively updated to obtain a context-aware target representation. The enhanced features are passed through the detection prediction head, and the parameters of the target category and the rotated bounding box are obtained by using the classification and regression branches, respectively.

[0044] Specifically, the extraction of multi-scale visual features from the original input image includes: A stripNet network pre-trained on a large image classification dataset is used as a visual encoder to extract multi-scale visual features from the input image.

[0045] Specifically, the context-enhanced modeling of multi-scale visual features involves: For each level of visual feature map, gridded pooling downsampling is performed to form a global context feature representation for the corresponding level. Based on this, the RoI (Region of Interest) features corresponding to the candidate target are obtained according to the feature level to which the candidate target belongs. The RoI features and the global context features of the same level are then input into the attention interaction module for feature interaction fusion to obtain an enhanced visual feature representation containing contextual information. Furthermore, residual fusion and feedforward transformation are performed on the enhanced visual feature representation to improve feature representation capability and training stability.

[0046] The method for constructing the hierarchical-specific global context representation is as follows: In the formula, The first output of the feature pyramid network Each level of feature map; This is a gridded adaptive average pooling operation used to... downsampling Grid features; This is an operation that unfolds according to a spatial grid; For the first A hierarchical global context representation, which includes Each context tag is used to represent the global semantic and spatial layout information of this level.

[0047] Furthermore, the method for obtaining enhanced visual feature representations containing contextual information through attention interaction is as follows: In the formula, RoI features for candidate target regions; RoI feature tokenization operation, used to... Mapped to RoI tag representation ; To correspond with the level to which the RoI belongs The corresponding hierarchical global context representation; For multi-head attention operations; For query, As key, Value; For context enhancement; The enhanced RoI tag representation after fusing contextual information, and the above It serves as an enhanced visual feature representation for subsequent candidate target region feature updates and detection predictions.

[0048] Specifically, the step of injecting the enhanced candidate target region labeling representation back into the candidate target region features involves: The candidate target region label representation, enhanced through attention interaction and global context, is returned to the original target region feature map. An adaptive update mechanism is then used to fuse the enhanced label representation with the original target region features, resulting in a more accurate and robust target representation, further improving the target features' ability to discriminate against complex background interference.

[0049] The method for adaptively updating the features of the candidate target region is as follows: In the formula, An enhanced RoI tag representation after incorporating contextual information; This is a normalization operation; It is a linear mapping used to generate modulation parameters; and These are the scale modulation term and the bias modulation term, respectively. RoI features for candidate target regions; This is element-wise multiplication; To modulate the updated structured enhanced RoI features, and to... This serves as the feature input for candidate target regions in subsequent detection and prediction.

[0050] Specifically, the process of passing the enhanced features through a detection prediction head and using classification and regression branches to obtain predictions of the target category and the rotated bounding box involves: The candidate target region features, after context enhancement and adaptive updating, are input into the detection prediction head. Combined with global context information, a collaborative feature representation for target category discrimination and rotation bounding box regression is obtained. Subsequently, the collaborative feature representation is input into the detection prediction head, wherein the classification branch outputs the target category confidence score, and the regression branch outputs the target's rotation bounding box parameters, including center point coordinates, width, height, and rotation angle.

[0051] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0052] The execution entities of the processor 1 and memory 2 mentioned above can be devices with computing functions such as computers, microcontrollers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution entity and can select it according to the needs of actual application.

[0053] Data signals are transmitted between memory 2 and processor 1 via bus 3, which will not be described in detail in this embodiment of the invention.

[0054] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0055] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0056] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0057] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0058] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0059] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for target detection in remote sensing images that resists interference from complex backgrounds, characterized in that, The method includes: Extract multi-scale visual features from the original input image; Context-enhanced modeling is performed on multi-scale visual features to construct hierarchical global contextual representations, and enhanced visual feature representations containing contextual information are obtained through attention interactions. The enhanced candidate target region label representation is injected back into the candidate target region feature, and the candidate target region feature is adaptively updated to obtain a context-aware target representation. The enhanced features are passed through the detection prediction head, and the parameters of the target category and the rotated bounding box are obtained by using the classification and regression branches, respectively.

2. The remote sensing image target detection method against complex background interference according to claim 1, characterized in that, The extraction of multi-scale visual features from the original input image specifically involves: A stripNet network pre-trained on a large image classification dataset is used as a visual encoder to extract multi-scale visual features from the input image.

3. The remote sensing image target detection method against complex background interference according to claim 1, characterized in that, The specific steps of performing context-enhanced modeling of multi-scale visual features are as follows: For each level of visual feature map, gridded pooling downsampling is performed to form a global context feature representation for the corresponding level. Based on this, the RoI (Region of Interest) features corresponding to the candidate target are obtained according to the feature level to which the candidate target belongs. The RoI features and the global context features of the same level are then input into the attention interaction module for feature interaction fusion to obtain an enhanced visual feature representation containing contextual information. Furthermore, residual fusion and feedforward transformation are performed on the enhanced visual feature representation to improve feature representation capability and training stability.

4. The remote sensing image target detection method against complex background interference according to claim 3, characterized in that, The method for constructing the hierarchical-specific global context representation is as follows: In the formula, The first output of the feature pyramid network Each level of feature map; This is a gridded adaptive average pooling operation used to... downsampling Grid features; This is an operation that unfolds according to a spatial grid; For the first A hierarchical global context representation, which includes Each context tag is used to represent the global semantic and spatial layout information of this level.

5. The remote sensing image target detection method against complex background interference according to claim 3, characterized in that, The method for obtaining enhanced visual feature representations containing contextual information through attention interaction is as follows: In the formula, RoI features for candidate target regions; RoI feature tokenization operation, used to... Mapped to RoI tag representation ; To correspond with the level to which the RoI belongs The corresponding hierarchical global context representation; For multi-head attention operations; For query, As key, Value; For context enhancement; The enhanced RoI tag representation after fusing contextual information, and the above It serves as an enhanced visual feature representation for subsequent candidate target region feature updates and detection predictions.

6. The remote sensing image target detection method against complex background interference according to claim 1, characterized in that, The specific steps of injecting the enhanced candidate target region labeling representation back into the candidate target region features are as follows: The candidate target region label representation, enhanced through attention interaction and global context, is returned to the original target region feature map. An adaptive update mechanism is then used to fuse the enhanced label representation with the original target region features, resulting in a more accurate and robust target representation, further improving the target features' ability to discriminate against complex background interference.

7. The remote sensing image target detection method against complex background interference according to claim 6, characterized in that, The method for adaptively updating the features of the candidate target region is as follows: In the formula, An enhanced RoI tag representation after incorporating contextual information; This is a normalization operation; It is a linear mapping used to generate modulation parameters; and These are the scale modulation term and the bias modulation term, respectively. RoI features for candidate target regions; This is element-wise multiplication; To modulate the updated structured enhanced RoI features, and to... This serves as the feature input for candidate target regions in subsequent detection and prediction.

8. The remote sensing image target detection method against complex background interference according to claim 1, characterized in that, The process of passing the enhanced features through a detection prediction head and using classification and regression branches to obtain predictions of the target category and the rotated bounding box is specifically as follows: The candidate target region features, after context enhancement and adaptive updating, are input into the detection prediction head. Combined with global context information, a collaborative feature representation is obtained for target category discrimination and bounding box regression. Subsequently, the collaborative feature representation is input into the detection prediction head, wherein the classification branch outputs the target category confidence, and the regression branch outputs the target's rotated bounding box parameters, including the center point coordinates, width, height, and rotation angle.

9. A remote sensing image target detection device resistant to complex background interference, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-8.

Images