Remote sensing image target detection method based on RT-DETR

Abstract

The invention discloses a remote sensing image target detection method based on RT-DETR, and the method comprises the following steps: 1, obtaining a remote sensing image data set, and completing the preprocessing; 2, inputting a backbone network, and extracting a multi-scale feature map; step 3, inputting a DSF module to realize multi-scale feature adaptive fusion; step 4, inputting an MSFE module, enhancing features and modeling long-range dependence; 5, embedding a G-HCO module in a corresponding stage of the backbone network, and optimizing feature expression; step 6, inputting an RT-DETR encoder-decoder, and outputting a target category and a target position; and 7, filtering the low-confidence prediction frame to obtain a final detection result. According to the remote sensing image target detection method based on the RT-DETR, through multi-scale feature adaptive fusion of the DSF module, frequency domain-spatial domain dual-domain feature enhancement of the MSFE module and global feature optimization of the G-HCO module, the method can also be expanded to multi-task scenes such as video target detection and segmentation-detection combination, and the generalization ability and the application range are better.

Description

Technical Field

[0001] This invention relates to the field of underwater geotechnical engineering construction technology, specifically to a remote sensing image target detection method based on RT-DETR. Background Technology

[0002] Remote sensing image target detection, as one of the core technologies of remote sensing data interpretation, is widely used in key fields such as military reconnaissance, agricultural resource monitoring, urban planning and construction, and environmental disaster assessment. With the rapid development of satellite remote sensing and aerial photography technologies, the resolution of remote sensing images has been continuously improved, and the types of identifiable targets have become increasingly rich. However, it also faces technical challenges brought about by complex scenes. Remote sensing images usually contain diverse background elements (such as the interweaving of urban buildings and roads, and the intermingling of farmland and water bodies), and targets exhibit characteristics such as large size differences (from small vehicles and building corners to large areas of water bodies and farmland), dense distribution (such as densely arranged residential buildings in urban areas and clustered factories in industrial parks), and irregular shapes (such as the winding boundaries of natural water bodies and roads in mountainous terrain). Traditional target detection methods such as Faster R-CNN and YOLO series have performed well in general image detection, but they are not adaptable to the special scenes of remote sensing images. In recent years, the DETR series models based on Transformer (such as RT-DETR) have achieved end-to-end detection, eliminating the traditional NMS post-processing steps, improving detection efficiency, and providing a new technical direction for remote sensing image target detection.

[0003] However, existing technologies still have significant shortcomings: RT-DETR has limited ability to retain details of small targets during feature extraction and fusion, making it difficult to handle scenarios where small targets account for a high proportion in remote sensing images; the YOLO series relies on the NMS algorithm, which can easily lead to missed detections of dense targets, and the non-end-to-end detection process increases the risk of error accumulation; traditional feature pyramid networks (such as FPN and PANet) are prone to losing local details of small targets when fusing high- and low-level features; at the same time, most existing methods do not fully utilize frequency domain information and long-range spatial dependencies, and lack the ability to perceive the global context of targets in complex backgrounds, ultimately affecting detection accuracy and robustness. Summary of the Invention

[0004] The purpose of this invention is to provide a remote sensing image target detection method based on RT-DETR, in order to solve the problems mentioned in the background art. Existing target detection models in remote sensing image detection suffer from low accuracy, easy missed detection and false detection, and inaccurate positioning due to complex backgrounds, small and densely distributed targets, and large differences in shape. They also suffer from poor feature fusion effect, insufficient global context awareness, and reliance on NMS post-processing.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a remote sensing image target detection method based on RT-DETR, comprising the following steps: Step 1: Obtain the remote sensing image dataset and complete the preprocessing; Step 2: Input the backbone network and extract multi-scale feature maps; Step 3: Input the DSF module to achieve adaptive fusion of multi-scale features; Step 4: Input the MSFE module to enhance features and model long-range dependencies; Step 5: Embed G-HCO modules in the corresponding stages of the backbone network to optimize feature representation; Step 6: Input RT-DETR encoder-decoder, output target category and location; Step 7: Filter out low-confidence prediction boxes to obtain the final detection results; In step one, a dataset of remote sensing images covering different scenes (such as urban buildings, farmland, water bodies, and roads), different resolutions, and different shooting conditions (such as lighting and weather) is first obtained through public remote sensing databases (such as Gaofen series satellite data and Google Earth Engine) or on-site aerial photography. Then, comprehensive preprocessing work is carried out, including data cleaning (deleting invalid images with severe blurring, noise interference, or missing / incorrect annotations), annotation verification and correction (adjusting offset target bounding boxes and supplementing missing small target annotations by comparing with the original scene information), image size normalization (scaling all images to a fixed size required by the model input to avoid size differences affecting training efficiency), and data augmentation operations (such as random rotation, horizontal / vertical flipping, adaptive adjustment of brightness and contrast, and adding reasonable noise). Through diversified data augmentation, the richness of the dataset is improved, enhancing the generalization ability and anti-interference ability of the subsequent model.

[0006] Preferably, in step two, the preprocessed remote sensing images are input in batches into a preset backbone network (such as ResNet, SwinTransformer, or other network structures adapted for remote sensing image feature extraction). The network generates multi-scale feature maps of different resolutions step by step through multi-level convolution operations (extracting local features), pooling operations (reducing feature map resolution and expanding the receptive field), and activation functions (introducing non-linear expressive power). Among them, the low-level feature maps have higher resolution and focus on capturing detailed information such as the edges and textures of the target in the image; the high-level feature maps have lower resolution but contain richer global semantic information (such as the overall shape and category features of the target), providing comprehensive feature support from details to the whole for subsequent target detection.

[0007] Preferably, in step three, all multi-scale feature maps output by the backbone network are input into the DSF (Multi-Scale Feature Adaptive Fusion) module. This module first unifies the dimensions and aligns the resolution of feature maps at different scales (by upsampling or downsampling operations to match the size of feature maps at each scale). Then, it introduces an adaptive weight calculation mechanism (such as an attention-based weight allocation algorithm) to dynamically adjust the contribution weight of each scale feature according to the feature importance of the target in different scenarios. For example, it assigns higher weights to low-level detail features that are key to small target detection and strengthens information fusion for high-level semantic features that are key to large target classification. Finally, the weighted multi-scale features are integrated to generate a fused feature map that combines detailed and global information, effectively solving the problem that single-scale features are insufficient to express targets of different sizes.

[0008] Preferably, in step four, the feature map fused by the DSF module is input into the MSFE (Feature Enhancement and Long-Range Dependency Modeling) module. Inside the module, the feature enhancement unit first enhances the target detail information in the fused features by using dilated convolution to expand the receptive field, introducing channel attention mechanism to highlight target feature channels, and using spatial attention mechanism to focus on target regions, thereby suppressing background noise interference. Then, by using long-range dependency modeling structures (such as the self-attention mechanism of Transformer and multi-layer dilated convolution stacking), the locality limitation of traditional convolution operations is broken, and the spatial and semantic associations between distant targets (such as scattered buildings and long roads) in remote sensing images are captured, further improving the feature map's ability to represent complex scenes (such as dense targets and occluded targets), and providing a more accurate feature basis for subsequent target recognition and localization.

[0009] Preferably, in step five, during the critical stage of feature extraction in the backbone network (the feature map output at this stage has both a certain resolution and rich semantic information, and is the core feature source for object detection), a G-HCO (Global Feature Optimization and Channel Interaction) module is embedded. This module first summarizes the global semantic information in the feature map through a global feature aggregation unit to avoid information fragmentation caused by local feature extraction. Then, through channel interaction mechanisms (such as channel shuffle and cross attention), it strengthens the information exchange between different feature channels, filters and retains feature channels that are effective for object detection, and suppresses the interference of redundant channels. At the same time, a spatial attention adjustment mechanism is introduced to focus on the areas in the feature map where the target may exist, further optimizing the accuracy of feature expression, so that subsequent modules can more efficiently utilize key features for object detection.

[0010] Preferably, in step six, the feature map optimized by the G-HCO module is input into the encoder-decoder structure of RT-DETR. The encoder part performs global feature association modeling on the input features through multiple Transformer encoder layers, mining the potential relationships between targets and between targets and the background in the feature map. At the same time, it performs dimensional compression and information refinement on the features to generate compact and semantically rich encoded features. The decoder part, based on preset anchor boxes or reference points, combines the global features output by the encoder and gradually generates prediction boxes for different targets through a cross-attention mechanism and a prediction head network. It also calculates the target category probability (such as "building", "water body", "vehicle" etc.) and precise location coordinates (such as the upper left and lower right corner coordinates of the target bounding box) corresponding to each prediction box, so as to achieve preliminary identification and localization of targets in remote sensing images.

[0011] Preferably, in step seven, firstly, based on the optimal confidence threshold determined during model training (usually determined through validation set performance evaluation, such as 0.5 or 0.6, balancing detection precision and recall), all predicted boxes output by the RT-DETR decoder are filtered out, removing predicted boxes with confidence levels below the threshold (these predicted boxes are mostly false detections or results with ambiguous localization); then, the non-maximum suppression (NMS) algorithm is used to deduplicate multiple overlapping predicted boxes within the same target area, retaining the predicted box with the highest confidence and most accurate localization; finally, the filtered prediction results are formatted and output as the final detection result containing the target category, bounding box coordinates, and corresponding confidence level, meeting the practical application needs of remote sensing image target detection (such as land surveying, environmental monitoring, urban planning, etc.).

[0012] Preferably, the parameters of the DSF module, MSFE module, and G-HCO module are optimized through joint training. The training process adopts an adaptive learning rate scheduling strategy (such as cosine annealing learning rate), and FocalLoss is combined to solve the imbalance problem of small target samples in remote sensing images. At the same time, IoULoss is introduced to optimize the bounding box regression accuracy. During model training, training, validation, and test sets are constructed based on public remote sensing datasets (VisDrone2019, RSOD). Overfitting is avoided through an early stopping mechanism, ensuring that the model can stably output high-precision detection results in remote sensing images of different scenarios.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: By using the multi-scale feature adaptive fusion of the DSF module, the frequency-spatial dual-domain feature enhancement of the MSFE module, and the global feature optimization of the G-HCO module, the detailed information of small targets is effectively preserved. This results in a 4.7% improvement in mAP@50 on the VisDrone2019 dataset and a 1.8% improvement on the RSOD dataset, significantly reducing false negatives and false positives for small targets in complex backgrounds, thus solving the core problem of insufficient representation of small targets in existing models. Simultaneously, by leveraging the long-range dependency modeling of the MSFE module and the global context interaction of the G-HCO module, the ability to identify dense targets and irregularly shaped targets is enhanced, while suppressing background noise interference. Compared with the YOLO series and other models that rely on NMS, this invention achieves significantly better performance. The original RT-DETR significantly improves the robustness of detection in remote sensing images under different resolutions and shooting conditions (lighting, weather). Furthermore, the core module adopts a lightweight design, enhancing feature representation while controlling computational complexity. The end-to-end architecture based on RT-DETR eliminates the NMS post-processing step, balancing detection accuracy and speed, and avoiding error accumulation in traditional non-end-to-end processes. In addition, through diverse data preprocessing and adaptive training strategies (FocalLoss, IoULoss, early stopping mechanism), the model can stably output high-precision results in remote sensing images of different scenes such as urban buildings, farmland, and water bodies. It can also be extended to multi-task scenarios such as video object detection and joint segmentation-detection, demonstrating superior generalization ability and application scope. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides a technical solution: a remote sensing image target detection method based on RT-DETR, comprising the following steps: Step 1: Obtain the remote sensing image dataset and complete the preprocessing; Step 2: Input the backbone network and extract multi-scale feature maps; Step 3: Input the DSF module to achieve adaptive fusion of multi-scale features; Step 4: Input the MSFE module to enhance features and model long-range dependencies; Step 5: Embed G-HCO modules in the corresponding stages of the backbone network to optimize feature representation; Step 6: Input RT-DETR encoder-decoder, output target category and location; Step 7: Filter out low-confidence prediction boxes to obtain the final detection results; In step one, a dataset of remote sensing images covering different scenes (such as urban buildings, farmland, water bodies, and roads), different resolutions, and different shooting conditions (such as lighting and weather) is first obtained through public remote sensing databases (such as Gaofen series satellite data and Google Earth Engine) or on-site aerial photography. Then, comprehensive preprocessing work is carried out, including data cleaning (deleting invalid images with severe blurring, noise interference, or missing / incorrect annotations), annotation verification and correction (adjusting offset target bounding boxes and supplementing missing small target annotations by comparing with the original scene information), image size normalization (scaling all images to a fixed size required by the model input to avoid size differences affecting training efficiency), and data augmentation operations (such as random rotation, horizontal / vertical flipping, adaptive adjustment of brightness and contrast, and adding reasonable noise). Through diversified data augmentation, the richness of the dataset is improved, enhancing the generalization ability and anti-interference ability of the subsequent model.

[0017] In step two, the preprocessed remote sensing images are input in batches into a preset backbone network (such as ResNet, SwinTransformer, or other network structures adapted for remote sensing image feature extraction). The network generates multi-scale feature maps of different resolutions step by step through multi-level convolution operations (extracting local features), pooling operations (reducing feature map resolution and expanding the receptive field), and activation functions (introducing non-linear expressive power). Among them, the low-level feature maps have higher resolution and focus on capturing detailed information such as the edges and textures of the target in the image; the high-level feature maps have lower resolution but contain richer global semantic information (such as the overall shape and category features of the target), providing comprehensive feature support from details to the whole for subsequent target detection.

[0018] In step three, all multi-scale feature maps output by the backbone network are input into the DSF (Adaptive Multi-Scale Feature Fusion) module. This module first unifies the dimensions and aligns the resolution of feature maps at different scales (by upsampling or downsampling operations to match the size of feature maps at each scale). Then, it introduces an adaptive weight calculation mechanism (such as an attention-based weight allocation algorithm) to dynamically adjust the contribution weight of each scale feature according to the feature importance of the target in different scenarios. For example, it assigns higher weights to low-level detail features that are key to small target detection and strengthens information fusion for high-level semantic features that are key to large target classification. Finally, the weighted multi-scale features are integrated to generate a fused feature map that combines detailed and global information, effectively solving the problem that single-scale features are insufficient to represent targets of different sizes.

[0019] In step four, the feature map fused by the DSF module is input into the MSFE (Feature Enhancement and Long-Range Dependency Modeling) module. Inside the module, the feature enhancement unit first strengthens the target detail information in the fused features and suppresses background noise interference by using dilated convolution to expand the receptive field, introducing channel attention mechanism to highlight target feature channels, and using spatial attention mechanism to focus on target regions. Then, with the help of long-range dependency modeling structure (such as the self-attention mechanism of Transformer and multi-layer dilated convolution stacking), the locality limitation of traditional convolution operation is broken, and the spatial and semantic associations between distant targets (such as scattered buildings and long roads) in remote sensing images are captured, which further improves the feature map's ability to represent complex scenes (such as dense targets and occluded targets) and provides a more accurate feature basis for subsequent target recognition and localization.

[0020] In step five, during the critical stage of feature extraction in the backbone network (the feature map output at this stage has both a certain resolution and rich semantic information, and is the core feature source for object detection), a G-HCO (Global Feature Optimization and Channel Interaction) module is embedded. This module first aggregates global semantic information in the feature map through a global feature aggregation unit to avoid information fragmentation caused by local feature extraction. Then, through channel interaction mechanisms (such as channel shuffle and cross attention), it strengthens the information exchange between different feature channels, filters and retains feature channels that are effective for object detection, and suppresses the interference of redundant channels. At the same time, a spatial attention adjustment mechanism is introduced to focus on the regions in the feature map where the target may exist, further optimizing the accuracy of feature representation, so that subsequent modules can more efficiently utilize key features for object detection.

[0021] In step six, the feature map optimized by the G-HCO module is input into the encoder-decoder structure of RT-DETR. The encoder part uses multiple Transformer encoder layers to perform global feature association modeling on the input features, mining the potential relationships between targets and between targets and the background in the feature map. At the same time, it performs dimensionality compression and information refinement on the features to generate compact and semantically rich encoded features. The decoder part, based on preset anchor boxes or reference points, combines the global features output by the encoder and uses a cross-attention mechanism and a prediction head network to gradually generate prediction boxes for different targets. It also calculates the target category probability (such as "building", "water body", "vehicle" etc.) and precise location coordinates (such as the upper left and lower right corner coordinates of the target bounding box) for each prediction box, realizing the preliminary identification and localization of targets in the remote sensing image.

[0022] In step seven, firstly, based on the optimal confidence threshold determined during model training (usually determined through validation set performance evaluation, such as 0.5 or 0.6, balancing detection precision and recall), all predicted boxes output by the RT-DETR decoder are filtered out, removing those with confidence scores below the threshold (these predicted boxes are mostly false detections or results with ambiguous localization). Then, the non-maximum suppression (NMS) algorithm is used to deduplicate multiple overlapping predicted boxes within the same target area, retaining the predicted box with the highest confidence and most accurate localization. Finally, the filtered prediction results are formatted and output as the final detection result, including the target category, bounding box coordinates, and corresponding confidence score, meeting the practical application needs of remote sensing image target detection (such as land surveying, environmental monitoring, urban planning, etc.).

[0023] The parameters of the DSF, MSFE, and G-HCO modules are optimized through joint training. The training process employs an adaptive learning rate scheduling strategy (such as cosine annealing), combined with FocalLoss to address the imbalance of small target samples in remote sensing images, and IoULoss to optimize bounding box regression accuracy. During model training, training, validation, and test sets are constructed based on publicly available remote sensing datasets (VisDrone2019, RSOD). An early stopping mechanism avoids overfitting, ensuring the model can consistently output high-precision detection results in remote sensing images across different scenarios. Through multi-scale feature adaptive fusion in the DSF module, frequency-spatial dual-domain feature enhancement in the MSFE module, and global feature optimization in the G-HCO module, detailed information about small targets is effectively preserved. On the VisDrone2019 dataset, mAP@50 is improved by 4.7%, and on the RSOD dataset by 1.8%, significantly reducing false negatives and false negatives of small targets in complex backgrounds, thus addressing the limitations of existing models in representing small targets. The core problem of insufficient features is addressed. Simultaneously, by leveraging the long-range dependency modeling of the MSFE module and the global context interaction of the G-HCO module, the model enhances the ability to identify dense targets and irregularly shaped targets while suppressing background noise interference. Compared to the YOLO series and the original RT-DETR, which rely on NMS, the robustness of detection in remote sensing images under different resolutions and shooting conditions (lighting, weather) is significantly improved. Furthermore, the core module adopts a lightweight design, controlling computational complexity while enhancing feature representation. The end-to-end architecture based on RT-DETR eliminates the NMS post-processing step, balancing detection accuracy and speed, and avoiding error accumulation in traditional non-end-to-end processes. In addition, through diverse data preprocessing and adaptive training strategies (FocalLoss, IoULoss, early stopping mechanism), the model can stably output high-precision results in remote sensing images of different scenes such as urban buildings, farmland, and water bodies. It can also be extended to multi-task scenarios such as video target detection and joint segmentation-detection, demonstrating superior generalization ability and application scope.

[0024] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A remote sensing image target detection method based on RT-DETR, characterized in that: The method comprises the following steps: Step one: obtain remote sensing image dataset, complete preprocessing; Step two: input the backbone network, extract multi-scale feature maps; Step three: input the DSF module, realize multi-scale feature adaptive fusion; Step four: input the MSFE module, enhance the feature and model the long-range dependence; Step five: embed the G-HCO module in the corresponding stage of the backbone network, optimize the feature expression; Step six: input the RT-DETR encoder-decoder, output the target category and position; Step seven: filter low confidence prediction boxes, get the final detection result; In step one, first, through public remote sensing database (such as high resolution satellite data, Google Earth Engine) or field aerial photography, remote sensing image dataset covering different scenes (such as urban buildings, farmland, water body, road), different resolution and different shooting conditions (such as light, weather) is obtained; Then carry out comprehensive preprocessing work, including data cleaning (delete invalid images with serious blur, noise interference or missing / incorrect label), label verification and correction (adjust the offset target bounding box according to the original scene information, supplement the missing small target label), image size normalization (unify all images to the fixed size required by the model input, avoid the influence of size difference on training efficiency), and data enhancement operation (such as random rotation, horizontal / vertical flip, brightness and contrast adaptive adjustment, adding reasonable noise, etc.), through diversified data expansion, improve the richness of the dataset, enhance the generalization ability and anti-interference ability of the subsequent model.

2. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: In step two, the preprocessed remote sensing image batch is input into the preset backbone network (such as ResNet, SwinTransformer, etc. Network structure suitable for remote sensing image feature extraction), the network generates multi-scale feature maps of different resolutions through multi-level convolution operation (extract local features), pooling operation (reduce feature map resolution, expand receptive field) and activation function (introduce nonlinear expression ability), including low-level feature map with high resolution, focusing on capturing the edge, texture and other detailed information of the target in the image; High-level feature map has lower resolution, but contains more global semantic information (such as target overall shape, class features), which provides comprehensive feature support from details to global for subsequent target detection.

3. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: In step three, the multi-scale feature maps output by the backbone network are all input into the DSF (multi-scale feature adaptive fusion) module. The module first unifies the dimensions and aligns the resolutions of the feature maps of different scales (through upsampling or downsampling operations to match the sizes of the feature maps of different scales), and then introduces an adaptive weight calculation mechanism (such as a weight distribution algorithm based on attention). According to the importance of the features of the target in different scenarios, the contribution weights of the features of different scales are dynamically adjusted. For example, higher weights are assigned to low-level detail features that are critical for small target detection, and high-level semantic features that are critical for large target classification are strengthened for information fusion. Finally, the weighted multi-scale features are integrated to generate a fusion feature map that combines detailed and global information, effectively solving the problem of insufficient expression of single-scale features for targets of different sizes.

4. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: In step four, the feature map fused by the DSF module is input into the MSFE (feature enhancement and long-range dependency modeling) module. Inside the module, the target detail information in the fused feature is first enhanced by a feature enhancement unit (such as using dilated convolution to expand the receptive field, introducing channel attention mechanism to highlight target feature channels, and using spatial attention mechanism to focus on target regions), and background noise interference is suppressed. Then, with the help of a long-range dependency modeling structure (such as the self-attention mechanism of Transformer and the stacking of multiple dilated convolutions), the locality limitation of traditional convolution operations is broken, and the spatial and semantic relationships between distant targets (such as scattered buildings and long roads) in remote sensing images are captured, further improving the feature map's representation ability for complex scenes (such as dense targets and occluded targets), and providing a more accurate feature basis for subsequent target recognition and positioning.

5. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: In step five, during the key stage of feature extraction by the backbone network (the feature map output at this stage has a certain resolution and rich semantic information, and is the core feature source for target detection), the G-HCO (global feature optimization and channel interaction) module is embedded. The module first aggregates global semantic information in the feature map through a global feature aggregation unit to avoid information fragmentation caused by local feature extraction. Then, through a channel interaction mechanism (such as channel shuffle and cross-attention), the information exchange between different feature channels is strengthened, and the effective feature channels for target detection are selected and retained, while the interference of redundant channels is suppressed. At the same time, a spatial attention adjustment mechanism is introduced to focus on the regions where the target may exist in the feature map, further optimizing the accuracy of feature expression, so that the subsequent modules can more efficiently utilize the key features for target detection.

6. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: In step six, the feature map optimized by the G-HCO module is input into the encoder-decoder structure of RT-DETR. The encoder part models the feature correlation in the global range through multiple Transformer encoder layers, excavates the potential relationships between targets and targets, and between targets and backgrounds in the feature map, and simultaneously compresses the dimensions and refines the information of the features to generate compact and semantically rich encoded features. The decoder part is based on the preset anchor box or reference point, combined with the global features output by the encoder, through the cross-attention mechanism and the prediction head network, to gradually generate prediction boxes for different targets, while calculating the target class probability (such as "building", "water", "vehicle", etc.) and accurate position coordinates (such as the top-left and bottom-right coordinates of the target bounding box) corresponding to each prediction box, realizing the preliminary identification and positioning of the target in the remote sensing image.

7. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: In step seven, first, according to the optimal confidence threshold determined in the model training process (usually determined by the performance evaluation of the validation set, such as 0.5 or 0.6, balancing the detection accuracy and recall rate), all the prediction boxes output by the RT-DETR decoder are screened, and the prediction boxes with confidence lower than the threshold are removed (such prediction boxes are mostly false positives or fuzzy positioning results); then the non-maximum suppression (NMS) algorithm is used to remove the multiple overlapping prediction boxes existing in the same target area, and the prediction box with the highest confidence and the most accurate positioning is retained; finally, the prediction results after screening are formatted, and the final detection results containing the target category, bounding box coordinates and corresponding confidence are output, meeting the actual application requirements of remote sensing image target detection (such as land surveying and mapping, environmental monitoring, urban planning, etc.).

8. The RT-DETR-based remote sensing image target detection method according to claim 1, characterized in that: The parameters of the DSF module, MSFE module and G-HCO module are optimized through joint training, and the adaptive learning rate scheduling strategy (such as cosine annealing learning rate) is used in the training process, combined with FocalLoss to solve the small target sample imbalance problem in remote sensing images, while introducing IoULoss to optimize the bounding box regression accuracy; during model training, the training set, validation set and test set are constructed based on the public remote sensing dataset (VisDrone2019, RSOD), the early stopping mechanism is used to avoid overfitting, and it is ensured that the model can stably output high-precision detection results in different scene remote sensing images.

Images