Remote sensing target detection method and system based on improved RT-DETR

Abstract

The invention discloses a remote sensing target detection method and system based on an improved RT-DETR, and the method comprises the following steps: 1, designing at least one HF-CSP module in a backbone network of the RT-DETR, and enabling the HF-CSP module to sense and extract the high-frequency features of a remote sensing target image; 2, introducing an adaptive sparse self-attention mechanism, and designing an ASSA-AIFI module in the RT-DETR for responding to the core region information of the remote sensing target image extracted by one HF-CSP module in the step 1; and step 3, designing a plurality of hybrid aggregation networks MANs in the RT-DETR for cross-scale feature fusion to obtain output features. The method shows higher detection performance and lower reasoning overhead, and is particularly suitable for remote sensing target detection application in real-time and resource limited scenes.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing target detection technology, specifically relating to a remote sensing target detection method and system based on an improved RT-DETR. Background Technology

[0002] Remote sensing target detection aims to automatically locate and identify targets on the ground using images acquired by multispectral or high-resolution sensors carried by satellites, aircraft, or UAVs. This technology has important application value in fields such as land resource surveys, urban planning, crop monitoring, disaster prevention and mitigation, and environmental protection (Liao Yurong, Wang Haining, Lin Cunbao, et al. Research progress on target detection of optical remote sensing images based on deep learning [J]. Journal of Communications, 2022, 43(05): 190-203.). Compared with natural scene images, remote sensing images have characteristics such as a top-down perspective, highly variable pixel scale, complex and diverse ground backgrounds, and small and sparsely or densely distributed targets. These characteristics make it difficult to accurately extract target information from remote sensing images (Zhang L, Xiong N, Gao W, et al. Improved detection method for micro-targets in remote sensing images [J]. Information, 2024, 15(2): 108.).

[0003] In recent years, with the continuous advancement of deep learning, convolutional neural networks (CNNs) have demonstrated excellent performance in image feature extraction and representation modeling, driving the continuous progress of object detection methods (Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Advances in neural information processing systems, 2012, 25.). Two-stage detectors, represented by Faster R-CNN (Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[J]. Advances in neural information processing systems, 2015, 28.), have a significant advantage in detection accuracy due to their strategy of separating candidate box generation and classification; while the YOLO series... Single-stage algorithms, on the other hand, are widely used in scenarios with high real-time requirements due to their efficiency.

[0004] Since then, Transformer-based visual models have attracted widespread attention in the field of object detection. Among them, the DetectionTransformer (DETR) proposed by the Facebook AI Research team in 2020 pioneered the introduction of the Transformer architecture into object detection tasks, breaking through the dependence of traditional methods on prior anchor boxes and post-processing strategies (such as non-maximum suppression), and realizing an end-to-end detection framework for the first time (Carion N, Massa F, Synnaeve G, etal. End-to-end object detection with transformers[C] / / European conference on computer vision. Cham: Springer International Publishing, 2020: 213-229.). Subsequently, to improve the convergence speed and computational efficiency of DETR, the Baidu team proposed RT-DETR (Real-Time Detection Transformer) in 2024. By optimizing the encoder-decoder structure, it effectively alleviated the problems of slow training and high resource consumption (Zhao Y, Lv W, Xu S, et al. Detrs beat yolos on real-time object detection[C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2024: 16965-16974.). However, since the self-attention mechanism tends to model global features, DETR-like models still have certain limitations in handling dense small targets or local information representation, especially in complex scenes containing a large number of small targets, where their local perception ability is relatively insufficient, resulting in limited detection accuracy and problems of false detection and missed detection. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention proposes a remote sensing image target detection method and system based on an improved RT-DETR architecture. Through three-stage structural optimization, the model's detection accuracy in complex scenes is significantly improved, while effectively reducing the number of parameters. First, in the backbone network, an HF-CSP module is designed, combining the ideas of HFERB and CSPNet, enhancing the model's ability to process high-frequency information, improving feature extraction accuracy, and significantly reducing model parameters. Second, an adaptive sparse self-attention mechanism is introduced to replace the multi-head self-attention structure in the original AIFI module, designing the ASSA-AIFI module, which effectively suppresses redundant computation and enhances the model's responsiveness to key information. Finally, in the cross-scale feature fusion stage, the original RepC3 module is replaced with a hybrid aggregation network, improving the diversity and semantic depth of feature representation.

[0006] The present invention adopts the following technical solution: A remote sensing target detection method based on improved RT-DETR, the specific steps of which are as follows: Step 1: In the backbone network of RT-DETR, design at least one HF-CSP module for sensing and extracting high-frequency features of remote sensing target images; Step 2: Introduce an adaptive sparse self-attention mechanism and design the ASSA-AIFI module in RT-DETR to respond to the core region information of the remote sensing target image extracted by one of the HF-CSP modules in Step 1. Step 3: Design multiple hybrid aggregation networks (MANs) in RT-DETR for cross-scale feature fusion to obtain output features.

[0007] In this invention, the HF-CSP module improves the ability to perceive and extract high-frequency features.

[0008] Preferably, the HF-CSP module introduces a high-frequency enhanced residual block HFERB and a cross-stage partial connection structure CSPNet; the input features are divided into two paths: one path directly performs cross-stage connection through the cross-stage partial connection structure; the other path inputs a stacked HFERB × n for feature enhancement; finally, the two paths are fused at the output. Here, n is a natural number.

[0009] Preferably, the high-frequency enhanced residual block (HFERB) is specifically as follows: Let the input feature map be... It is divided into two branches in terms of channel dimension: in, H represents the input to the local feature extraction branch and the high-frequency enhancement branch, respectively; H represents the height of the feature map; W represents the width of the feature map; and C represents the number of channels in the feature map. Local branches are extracted using standard 3×3 convolution and GELU activation to extract basic texture features: The high-frequency branch captures high-frequency information through max pooling and enhances the feature response using 1×1 convolution and GELU activation. The two feature streams are concatenated and then fused using a 1×1 convolution: The final residual output is obtained as follows: In this invention, the ASSA-AIFI module employs an adaptive sparse self-attention mechanism, which enhances the model's response to information in the core region.

[0010] Preferably, the ASSA-AIFI module constructs two parallel attention branches: a sparse attention branch (SSA) and a dense attention branch (DSA), and adaptively weights and fuses the two expressions to achieve the goal of balancing information selectivity and expression integrity.

[0011] Preferably, the ASSA-AIFI module is as follows: For the input feature map First, it is divided into multiple non-overlapping local windows of size M×M; each window is unfolded into a two-dimensional matrix. For each feature within a window, a uniform linear transformation is used to construct a vector of query Q, key K, and value V, as shown below: in, The weight matrix is ​​shared, corresponding to the query, key, and value projection operations, respectively; Within each window, the weight matrices of the dense attention branch DSA and the sparse attention branch SSA are computed in parallel. Dense attention branch DSA normalizes all QK pairs using the standard Softmax function to compute the attention response: Where d represents the dimensions of Q and K, and B represents a trainable relative position bias term used to encode spatial position information; The Sparse Attention Branch (SSA) introduces the squared ReLU activation function: ASSA introduces an adaptive fusion mechanism, using learnable weight parameters. Perform a linear combination of the outputs of the two branches: Where e is the natural constant, These are learnable parameters.

[0012] In this invention, the hybrid aggregation network MAN can improve feature diversity and semantic depth.

[0013] Preferably, the MAN is as follows: For each layer of feature maps to be fused The number of channels is increased to 2C through a 1×1 convolution to generate intermediate features. It is divided into three parallel sub-paths: The first path is to X mid By directly inputting a single-layer 1×1 convolution, the number of channels is reduced from 2C to C, enabling channel-level information filtering and mapping. For the feature extraction requirements in the spatial dimension, the second path is to extract the intermediate feature X. mid The input depth can be processed using the separable convolution DSConv. The third path will use the intermediate feature X mid The features are split into two sub-features, X3 and X4, which are used as the initial inputs in the C2f module. A deep residual path is constructed by stacking multiple convolutional neck modules (ConvNeck), and each convolutional unit uses a skip connection. Finally, the features generated by the three paths are concatenated along the channel dimension and compressed to the original number of channels through a 1×1 convolution to obtain the final output features; This invention also discloses a remote sensing target detection system based on an improved RT-DETR, used to perform the above method, specifically including the following modules: HF-CSP module: Located in the backbone network of RT-DETR, it is used for sensing and extracting high-frequency features of remote sensing target images; ASSA-AIFI module: Located in RT-DETR, used to respond to information about the core region of remotely sensed target images; Hybrid Aggregation Network (MAN): Located in RT-DETR, it is used for cross-scale feature fusion to obtain output features.

[0014] This invention proposes a remote sensing target detection method and system based on an improved RT-DETR. First, an HF-CSP module is designed in the backbone to enhance the model's ability to perceive high-frequency information such as edges and textures, while significantly reducing redundant gradient flows and the number of parameters. Next, an adaptive sparse self-attention mechanism is introduced into the AIFI module, which can dynamically balance sparsity and integrity, effectively focusing on key regions in remote sensing images and suppressing background interference. Finally, in the original CCFF module, a hybrid aggregation network replaces the traditional RepC3, enabling multi-path parallel extraction of channel, spatial, and deep semantic features, enhancing the diversity and semantic depth of feature representation. Attached Figure Description

[0015] Figure 1 This is a structural diagram of the improved RT-DETR according to a preferred embodiment of the present invention.

[0016] Figure 2 This is a structural diagram of the HFERB according to a preferred embodiment of the present invention.

[0017] Figure 3 This is a structural diagram of the HF-CSP module according to a preferred embodiment of the present invention.

[0018] Figure 4 This is a structural diagram of the ASSA-AIFI module according to a preferred embodiment of the present invention.

[0019] Figure 5 This is a structural diagram of the MAN module of a preferred embodiment of the present invention.

[0020] Figure 6 This is a comparison chart of the test results. Detailed Implementation

[0021] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] This embodiment presents a remote sensing target detection method based on an improved RT-DETR, the specific steps of which are as follows: Step 1: In the backbone network of RT-DETR, design at least one HF-CSP module combining the ideas of HFERB and CSPNet to perceive and extract high-frequency features of remote sensing target images; the extracted high-frequency features serve as the input basis for subsequent feature enhancement and aggregation. Step 2: Introduce an adaptive sparse self-attention mechanism to replace the multi-head self-attention structure in the original AIFI module. Design an ASSA-AIFI module in RT-DETR to respond to and model the core region information of the high-frequency features of the remote sensing target image extracted by one of the HF-CSP modules in Step 1. The obtained core region response features are used to enhance the effectiveness of cross-scale feature fusion. Step 3: In the cross-scale feature fusion stage, the original RepC3 module is replaced with the hybrid aggregation network MAN, which is used to perform cross-scale feature fusion of the high-frequency features extracted in step 1 and the core region response features obtained in step 2 to obtain the output features.

[0023] The following is a more detailed description of this embodiment.

[0024] Introduction to the RT-RETR model: RT-DETR is an end-to-end object detection model based on the Transformer architecture, designed to eliminate the reliance on non-maximum suppression in traditional detection methods, thereby reducing inference latency and improving detection efficiency. This model achieves overall optimization of object localization and classification through a unified encoder-decoder structure.

[0025] The model mainly consists of three parts: a backbone network, an efficient hybrid encoder, and a decoder. The image is first processed by a backbone network (such as ResNet) to extract multi-scale features (S3, S4, S5), which are then fed into the encoder for further processing. The encoder consists of two core modules: the AIFI (Attention-based Intra-scale Feature Interaction) module uses a Transformer to perform feature interaction within a single scale on the S5 features to enhance the modeling ability for high semantic information; the CCFF (CNN-based Cross-scale Feature Fusion) module fuses features from different scales through convolution, improving the richness and stability of the overall feature representation. In the decoding stage, the model generates an initial query through an IOU-aware mechanism, selecting the most representative feature locations for target prediction. Subsequently, the decoder refines these queries through multi-layer interactions, outputting the final bounding box and class label. An introduction to the improved RT-DETR model of the preferred embodiment of this invention is provided below. To address the false positives and false negatives that occur in existing RT-DETR models for remote sensing target detection, this invention proposes improvements in three areas. The improved model structure is as follows: Figure 1 As shown in the diagram. In the backbone network, this embodiment designs a new module, HF-CSP, which effectively improves the model's ability to perceive and extract high-frequency features, while significantly reducing the overall number of parameters in terms of structural lightweighting. Subsequently, in the attention mechanism, the original AIFI module's multi-head self-attention is replaced with an adaptive sparse self-attention mechanism, enhancing the model's response to information in core regions. Finally, in the feature fusion stage, a hybrid aggregation network is introduced to replace the traditional RepC3 module, which can capture richer semantic levels and diverse structural features, thereby comprehensively improving the overall performance of remote sensing target detection. These three modules are described in detail below.

[0026] 1. HF-CSP module Remote sensing images are characterized by their wide spatial coverage and diverse types of ground features. They are often accompanied by problems such as large scale spans and strong background interference, which places higher demands on the feature extraction capabilities of target detection models, especially their ability to perceive high-frequency information. To address this, this invention introduces a High-Frequency Enhanced Residual Block (HFERB) (Li A, Zhang L, Liu Y, et al. Feature modulation transformer: Cross-refinement of global representation via high-frequency prior for image super-resolution[C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2023: 12514-12524.) and a CrossStage Partial Network (CSPNet) (Wang CY, Liao HYM, Wu YH, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition workshops. 2020:390-391.) into the RT-DETR backbone network, designing an improved module HF-CSP that integrates high-frequency enhancement and structural optimization to improve the model's target detection performance in complex scenes. Improved HF-CSP module structure, such as Figure 3 As shown.

[0027] like Figure 2 As shown, HFERB borrows from frequency domain concepts, emphasizing the response to high-frequency information in images (such as edges and textures). Specifically, let the input feature map (i.e., the feature map of the HFERB module) be... First, it is divided into two branches in terms of channel dimension: in, These represent the inputs to the local feature extraction branch and the high-frequency enhancement branch, respectively.

[0028] Local branches are extracted using standard 3×3 convolution and GELU activation to extract basic texture features: The high-frequency branch captures high-frequency information through max pooling, followed by 1×1 convolution and GELU activation to enhance the feature response. The two feature streams are concatenated and then fused using a 1×1 convolution: The final residual output is obtained as follows: Inspired by CSPNet, this invention introduces feature splitting and fusion strategies into the module structure to reduce gradient information redundancy and optimize computational overhead, designing the HF-CSP module, the structure of which is as follows: Figure 3 As shown in the diagram. Specifically, the input features are divided into two paths: one path directly performs cross-stage connections to maintain the integrity of shallow features; the other path inputs stacked deep feature extraction units (HFERB × n) for feature enhancement. Finally, the two paths are fused at the module output to achieve a balanced distribution of gradient flow, alleviate overfitting, and improve parameter utilization. This design significantly reduces redundant gradient backflow and invalid computational overhead while ensuring feature representation capabilities. Furthermore, the fused features retain both shallow detail information and deep semantic features, enabling the model to accurately locate and identify targets even in complex backgrounds, demonstrating stronger robustness and discriminative ability.

[0029] like Figure 1 As shown, multiple HF-CSPs are used in the backbone network. Input is the initial input, which is then processed by multiple HF-CSPs. The outputs of layers S3 and S4 enter the cross-scale feature fusion, and the output of the last layer (S5) is input to the ASSA-AIFI module.

[0030] 2. ASSA-AIFI In RT-DETR, the AIFI module employs a multi-head self-attention mechanism (Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural informationprocessing systems, 2017, 30.) to achieve information interaction between features. However, its original design, based on the dense self-attention structure of the standard Transformer, does not distinguish the importance of regions when calculating the attention between all tokens. This indiscriminate interaction method is easily affected by a large number of non-target regions when processing remote sensing images, especially in high-resolution images with a large number of dense small targets. The model's region of interest is easily dispersed, leading to redundant feature representations and thus affecting detection accuracy.

[0031] To address this, this invention introduces an Adaptive Sparse Self-Attention (ASSA) mechanism (Zhou S, Chen D, Pan J, et al. Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration[C] / / Proceedings of the IEEE / CVFconference on computer vision and pattern recognition. 2024: 2952-2963.) to replace the original multi-head self-attention mechanism in the AIFI module. This explicitly guides attention to focus on semantically relevant regions and suppresses invalid responses. The module structure is as follows: Figure 4 As shown.

[0032] ASSA's design is based on a key observation: in real-world images, truly meaningful feature regions are typically sparsely distributed, especially in remote sensing scenes where background regions, although occupying most pixels, have extremely low information density. Therefore, ASSA introduces a sparse modeling strategy into its self-attention mechanism to enhance region selectivity. ASSA simultaneously constructs two parallel attention branches: a sparse branch (SSA) and a dense branch (DSA), and achieves a balance between information selectivity and representational integrity through adaptive weighted fusion of these two representations. The specific process is as follows: The input feature map (the feature map output from the last layer (S5) of the backbone network, which has undergone feature extraction by the HF-CSP module in multiple steps 1) First, it is divided into multiple non-overlapping local windows of size M x M. Each window is unfolded into a two-dimensional matrix. This is to facilitate subsequent attention modeling. For each feature within a window, a uniform linear transformation is used to construct query (Q), key (K), and value (V) vectors, as shown below: in, The weight matrix is ​​shared and corresponds to the projection operations of the query, key, and value, respectively.

[0033] Next, the weight matrices of the dense attention branch (DSA) and sparse attention branch (SSA) are computed in parallel within each window (see Equations 7 and 8).

[0034] The Dense Attention Branch (DSA) calculates its attention response by normalizing all QK pairs using the standard Softmax function: Here, B represents a trainable relative position bias term used to encode spatial position information.

[0035] The Sparse Attention Branch (SSA) introduces a squared ReLU activation function to suppress negatively correlated regions, retaining only the parts that have a positive correlation with the current query, thereby improving the sparsity and discriminativeness of the attention distribution. While SSA can significantly reduce interference from irrelevant regions, its excessive sparsity can lead to the loss of crucial information. To achieve a dynamic balance between sparsity and representational integrity, ASSA introduces an adaptive fusion mechanism that linearly combines the outputs of the two branches using learnable weight parameters w1 and w2: in, This is a learnable parameter, initially set to 1. This mechanism can dynamically control the sparsity of attention based on image feature content, thereby balancing the trade-off between filtering redundancy and preserving information.

[0036] This invention replaces the original attention mechanism in AIFI with the ASSA module while keeping other structures unchanged, thereby alleviating problems such as small targets being easily submerged by the background and attention being diverted in remote sensing scenes, and enhancing the discriminative representation ability of target features.

[0037] 3. MAN While the traditional RepC3 module offers advantages in terms of lightweight design, its relatively simple structure makes it difficult to fully capture the complex spatial patterns and semantic relationships in remote sensing images. Therefore, to further enhance the feature modeling capabilities of RT-DETR in remote sensing target detection, this invention optimizes the structure of its cross-scale feature fusion module (CCFF), replacing the original RepC3 module with a Mixed Aggregation Network (MAN) (Feng Y, Huang J, Du S, et al. Hyper-yolo: When visualobject detection meets hypergraphcomputation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.), thereby improving feature diversity and semantic depth. The structure of MAN is as follows: Figure 5 As shown. Detailed explanation follows: First, for each layer of feature maps to be fused (e.g.) Figure 1 As shown, there are multiple MANs here, and the feature maps of each input are different. Some are concatenated from S5 of step 1 and S2, while others are concatenated from the output of the previous MAN and S3 of step 1. That is, it will fuse the outputs from different layers of the backbone network (the outputs after HF-CSP and ASSA-AIFI). The number of channels is increased to 2C through a 1×1 convolution to generate intermediate features. Subsequently, X mid It is divided into three parallel sub-paths: The first path is to X mid By directly inputting a single-layer 1×1 convolution, the number of channels is reduced from 2C to C, enabling channel-level information filtering and mapping, which helps to enhance the network's responsiveness to channels related to the target category.

[0038] Meanwhile, for the feature extraction requirements in the spatial dimension, the second path is to extract the intermediate feature X. midThe input is processed using depthwise separable convolution (DSConv) (Howard AG, Zhu M, Chen B, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv preprint arXiv:1704.04861, 2017.), a step that can effectively capture local structural information in space.

[0039] Subsequently, in order to improve the semantic hierarchical integration capability of features, the third path integrates the intermediate feature X. mid The features are split into two sub-features, X3 and X4, which serve as the initial inputs to the C2f module. A deep residual path is constructed by stacking multiple convolutional neck modules (ConvNeck), with each convolutional unit employing skip connections to enhance information flow. This path is used to model deeper local features and enhances gradient flow through residual connections, avoiding information degradation and convergence difficulties.

[0040] Finally, the features generated by the three paths are concatenated along the channel dimension and compressed to the original number of channels using a 1×1 convolution to obtain the final output features.

[0041] In this embodiment, RepC3 in the CCFF module is completely replaced with MAN, which can effectively improve the diversity and semantic depth of feature expression, and realize deeper semantic fusion and broader feature coverage.

[0042] The following experiments demonstrate the significant advantages of this invention.

[0043] 1. Dataset The dataset used in this invention is the DIOR dataset (Li K, Wan G, Cheng G, et al. Objectdetection in optical remote sensing images: A survey and a new benchmark[J].ISPRS journal of photogrammetry and remote sensing, 2020, 159: 296-307.). The DIOR dataset is a large-scale object detection benchmark dataset for optical remote sensing images, widely used to evaluate the generalization ability and robustness of object detection models in complex remote sensing scenes. This dataset contains 23,463 remote sensing images, covering 192,472 annotated object instances, encompassing 20 typical land cover categories.

[0044] In this experiment, the dataset was split using the official original partition, consisting of 5862 training images, 5863 validation images, and 11738 test images.

[0045] 2. Experimental Environment The experimental environment for this invention was an Ubuntu 18.04 operating system, an NVIDIA A40 48GB GPU, and Python version 3.8.16, PyTorch version 1.13.1, and CUDA version 11.7.

[0046] In the experiment, the hyperparameter settings for model training were as follows: 300 training epochs, batch size of 4, and input image size of 640×640.

[0047] 3. Evaluation Indicators In the experiment, several commonly used metrics were used to evaluate the model performance, including the number of parameters, floating-point operations (GFLOPs), accuracy (P), recall (R), and mean average precision (mAP).

[0048] Accuracy P measures the proportion of samples that the model correctly identifies as positive out of all samples predicted as positive, reflecting the precision of its classification results. It is defined as follows: Where TP stands for True Positives, representing the number of samples correctly predicted as positive by the model, and FP stands for False Positives, representing the number of negative samples that are incorrectly classified as positive.

[0049] Recall R describes the coverage of the model with true positive samples, that is, the proportion of samples that are correctly identified out of all actual positive samples. The formula is as follows: FN stands for False Negatives, which is the number of positive class samples that the model failed to identify.

[0050] Mean accuracy (mAP) is a core metric for evaluating overall performance in object detection tasks. First, the average accuracy (AP) for each category is calculated. Then, the average AP across all categories is taken to obtain mAP. The specific definition is as follows: Where n is the number of categories, AP i This represents the area under the Precision-Recall curve for the i-th class at different recall rates. AP is defined as follows: In practical evaluations, mAP is typically calculated using different IoU thresholds to reflect the detector's performance under varying positioning accuracy requirements. Two thresholds are chosen: one is mAP50, representing the average accuracy at an IoU threshold of 0.5, which is a more lenient evaluation standard; the other is mAP50-95, representing the average mAP across 10 thresholds ranging from 0.5 to 0.95 (with a step size of 0.05), comprehensively reflecting the model's detection capability under different levels of overlap, and thus providing a more stringent evaluation standard.

[0051] 4. Experimental Results 4.1 Module Ablation Experiment surface Module ablation experiment To verify the effectiveness of the three proposed modules, multiple module ablation experiments were conducted on the DIOR public dataset, using the RT-DETR model as the baseline model. The experimental results are shown in Table 1. The final experimental results were evaluated using precision, recall, mean precision (mAP50, mAP50-95), number of parameters, and floating-point computation cost. The ablation experiment results are shown in the table.

[0052] Experimental results show that, firstly, after introducing the CSP-HF module into the baseline model RT-DETR, the detection accuracy is slightly improved by 0.2% for mAP50 and 0.6% for mAP50-95. However, the overall computational cost of the model is significantly reduced from 57.0 GFLOPs to 43.6 GFLOPs, and the number of parameters is also reduced from 19.9M to 13.0M, demonstrating the effectiveness of this module in reducing model complexity and improving inference efficiency.

[0053] Secondly, when the ASSA-AIFI module was introduced alone, the model's accuracy performance was significantly improved, with precision, recall, mAP50, and mAP50-95 increasing by 0.7%, 1.7%, 1.5%, and 0.5%, respectively. This demonstrates the effectiveness of the ASSA-AIFI module in enhancing the discriminative representation of target features and mitigating attentional distraction. Building on this, combining the ASSA-AIFI and MAN modules, despite increasing the number of parameters and floating-point computations to 22.8M and 60.2 GFLOPs, respectively, resulted in the most significant performance improvement. Accuracy, recall, mAP50, and mAP50-95 were improved by 1.2%, 1.9%, 1.9%, and 1.5%, respectively, indicating a good synergistic effect between the two modules in feature enhancement and feature fusion.

[0054] Ultimately, by combining all three modules, the model achieves a good balance between overall performance and efficiency, with mAP50 reaching 68.1% and mAP50-95 at 47.9%, representing improvements of 1.9% and 1.4% respectively compared to the original model. The improved model has a parameter count of 15.9M and GFLOPs of 46.8, representing reductions of 20% and 18% respectively compared to the original model. These results demonstrate that the improved model can achieve both high performance and lightweight design in remote sensing target detection tasks.

[0055] 4.2 Comparative Experiment surface Improved model comparison experiment To further verify the effectiveness of the improved model proposed in this invention, its performance was compared with that of existing mainstream detectors on the DIOR dataset. The comparison results are shown in Table 2. The comparison models include YOLOv5s, YOLOv8s, and YOLOv10m from the YOLO series.

[0056] Experimental results show that the improved RT-DETR proposed in this invention outperforms the comparative models while maintaining a moderate number of parameters and computational complexity. Specifically, our model achieves precision of 81.8% and recall of 62.8%, which are superior to RT-DETR-R18's 81.3% and 61.5%, respectively, indicating higher accuracy and fewer false negatives and missed detections. Compared with YOLOv10m, which has a similar number of parameters, our improved model improves mAP50 and mAP50-95 by 0.3% and 0.4%, respectively, while reducing computational complexity by 16.7 GFLOPs, further validating its higher inference efficiency and deployment advantages while maintaining detection accuracy.

[0057] 4.3 Visual Result Analysis Figure 6 This paper presents a comparison of the detection performance of the original RT-DETR model and the improved RT-DETR model of this invention on the DIOR test set. From... Figure 6 The comparison between (a) and (b) shows that the target performance of the improved RT-DETR in this invention is significantly enhanced. Taking the wind turbine on the left as an example, the original RT-DETR only detected two scattered targets in the image, and the confidence scores were generally low; while the improved model in this invention not only detected other wind turbines in the image, but also generally improved the confidence scores to the range of 0.40–0.69, indicating that it has a stronger perceptual ability in small target detection and weak feature representation.

[0058] In the image on the right, the improved model of this invention also performs excellently in detecting highway toll stations and vehicles. Although the original model can detect toll stations well, it misses a large number of vehicle targets in the image. The improved model, while maintaining the accuracy of toll station detection, also identifies two additional vehicles, which significantly improves the model's ability to recall multi-class targets in complex backgrounds.

[0059] Experimental results show that the improved RT-DETR model of this invention has stronger target perception capability, higher detection accuracy and better small target detection effect than the original RT-DETR in remote sensing target detection scenarios, verifying the effectiveness of the proposed improved method.

[0060] A preferred embodiment of the present invention also discloses a remote sensing target detection system based on an improved RT-DETR, used to perform the above method, specifically including the following modules: HF-CSP module: Located in the backbone network of RT-DETR, it is used for sensing and extracting high-frequency features of remote sensing target images; ASSA-AIFI module: Located in RT-DETR, used to respond to information about the core region of remotely sensed target images; Hybrid Aggregation Network (MAN): Located in RT-DETR, it is used for cross-scale feature fusion to obtain output features.

[0061] Other aspects of this embodiment can be found in the above method embodiments.

[0062] In summary, to address the issues of false detection and missed detection in remote sensing target detection, this invention proposes an improved target detection algorithm based on RT-DETR. First, an HF-CSP module is designed in the backbone to enhance the model's ability to perceive high-frequency information such as edges and textures, while significantly reducing redundant gradient flows and the number of parameters. Next, an adaptive sparse self-attention mechanism is introduced into the AIFI module, which can dynamically balance sparsity and integrity, effectively focusing on key areas in remote sensing images and suppressing background interference. Finally, in the original CCFF module, a hybrid aggregation network replaces the traditional RepC3, enabling multi-path parallel extraction of channel, spatial, and deep semantic features, enhancing the diversity and semantic depth of feature representation.

[0063] A series of experimental results on the DIOR dataset demonstrate that the improved model of this invention achieves a good balance between detection accuracy, efficiency, and lightweight design. Compared with mainstream methods such as the YOLO series, this invention also exhibits higher detection performance and lower inference overhead, making it particularly suitable for remote sensing target detection applications in real-time and resource-constrained scenarios.

[0064] The preferred embodiments of the present invention have been described in detail above. It will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An improved RT-DETR-based remote sensing target detection method, characterized in that, The specific steps are as follows: Step 1: In the backbone network of RT-DETR, at least one HF-CSP module is designed for extracting high-frequency features of remote sensing target images; Step 2: An adaptive sparse self-attention mechanism is introduced, and an ASSA-AIFI module is designed in RT-DETR to respond to the core area information of the remote sensing target image extracted by one of the HF-CSP modules in step 1; Step 3: Multiple mixed aggregation networks MAN are designed in RT-DETR for cross-scale feature fusion to obtain output features.

2. The remote sensing target detection method based on improved RT-DETR according to claim 1, characterized in that, The HF-CSP module introduces a high-frequency enhancement residual block HFERB and a cross-stage partial connection structure CSPNet; the input features are divided into two paths: one path directly connects through the cross-stage partial connection structure; the other path inputs the stacked HFERB x n for feature enhancement; Finally, the two paths of features are fused at the output end.

3. The remote sensing target detection method based on improved RT-DETR according to claim 2, characterized in that, The high-frequency enhancement residual block HFERB is specifically as follows: assuming that an input feature map is is divided into two branches in the channel dimension: wherein, respectively represent the input of the local feature extraction branch and the high-frequency enhancement branch, H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels of the feature map; The local branch uses standard 3x3 convolution and GELU activation to extract basic texture features: The high-frequency branch captures high-frequency information through max pooling and uses 1x1 convolution and GELU activation to enhance feature response: The two paths of features are spliced and fused through 1x1 convolution: Finally, the residual output is obtained: 。 4. The remote sensing target detection method based on improved RT-DETR of claim 1, characterized in that, The ASSA-AIFI module constructs two parallel attention branches: a sparse attention branch SSA and a dense attention branch DSA, and fuses the two expressions through adaptive weighting to achieve the goal of balancing information selectivity and expression integrity.

5. The remote sensing target detection method based on improved RT-DETR according to claim 4, characterized in that, The ASSA-AIFI module is as follows: For an input feature map , it is first segmented into a plurality of non-overlapping local windows of size M x M; each window is unfolded into a two-dimensional matrix ; for each feature within a window, a unified linear transformation is adopted to construct a query Q, a key K, and a value V vector, represented as follows: wherein, are shared weight matrices, respectively corresponding to the projection operations of the query, key and value; The weight matrix of the dense attention branch DSA and the sparse attention branch SSA is calculated in parallel within each window; The dense attention branch DSA normalizes all Q-K pairs through the standard Softmax function to calculate the attention response: Where d represents the dimension of Q and K, and B represents a trainable relative position bias term used to encode spatial position information; The sparse attention branch SSA introduces a square ReLU activation function: The ASSA linearly combines the outputs of the two branches through learnable weight parameters w1 and w2: where e is a natural constant, are learnable parameters.

6. The remote sensing target detection method based on improved RT-DETR of claim 1, wherein, The MAN is as follows: For each layer of the feature map to be fused The number of channels is expanded to 2C by a 1x1 convolution to generate intermediate features Is divided into three parallel sub-paths: The first path is to combine X mid Directly input a single layer 1x1 convolution to reduce the channel number from 2C to C, realize channel-level information filtering and mapping; For the feature extraction requirement of spatial dimension, the second path is to extract the intermediate feature X mid is input into a depth separable convolution DSConv for processing; The third path splits the intermediate feature X mid into two sub-features X3 and X4 as initial inputs in C2f modules; a deep residual path is constructed by stacking multiple convolution neck modules ConvNeck, and each convolution unit adopts a skip connection; Finally, the features generated by the three paths are spliced in the channel dimension, and a 1x1 convolution is used to compress to the original channel number to obtain the final output features; 。 7. An improved RT-DETR based remote sensing target detection system for performing the method of any one of claims 1-6, characterized in that, Specifically, the following modules are included: HF-CSP module: located in the backbone network of RT-DETR, used for sensing and extracting high-frequency features of remote sensing target images; ASSA-AIFI module: located in RT-DETR, used for responding to core area information of remote sensing target images; Mixed aggregation network MAN: located in RT-DETR, used for cross-scale feature fusion to obtain output features.

Images