A sparse-gated dual-domain target detection method for UAV images

Abstract

This invention discloses a sparse-gated dual-domain target detection method for UAV images, comprising the following steps: Step S1: Acquire UAV image data and preprocess the UAV image data to obtain preprocessed image data; Step S2: Design a sparse-gated dual-domain target detection model; Step S3: Train the sparse-gated dual-domain target detection model using the training set in the preprocessed image data, and obtain the optimal target detection model through iterative optimization; Step S4: Use the optimal target detection model to perform target detection on the acquired UAV images. This sparse-gated dual-domain target detection method for UAV images achieves synergistic optimization in three dimensions: detection accuracy, model lightweighting, and inference efficiency, effectively overcoming the limitations of existing technologies in terms of computational efficiency, small target detection, and environmental adaptability.

Description

Technical Field

[0001] This invention relates to the field of target detection in UAV images, and specifically to a sparse gated dual-domain target detection method for UAV images. Background Technology

[0002] The deep integration of deep learning and edge computing enables unmanned aerial vehicle (UAV) systems to deploy advanced artificial intelligence algorithms, thereby gaining autonomous capabilities for real-time perception and decision-making.

[0003] While drones offer a unique aerial perspective for object detection, the images they capture present numerous unique challenges. Pedestrians, vehicles, and other targets in these images typically have a low pixel ratio, with approximately 60% to 80% of objects smaller than 32×32 pixels, resulting in sparse target features and limited recognizability. In complex urban or natural backgrounds, severe occlusion between targets further complicates feature extraction and classification. Simultaneously, dynamic changes in drone flight altitude and shooting angle cause frequent fluctuations in image scale and insufficient image stability. Adverse weather conditions such as rain, snow, and fog also significantly degrade image quality, affecting recognition accuracy. More importantly, drone-borne computing platforms are generally limited by energy consumption, computing power, and storage space, requiring edge-deployed detection algorithms to possess high accuracy, low parameter count, and high computational efficiency to adapt to demanding hardware conditions. Traditional detection methods often struggle to address these challenges simultaneously. For example, filter-based methods perform reasonably well in uniform backgrounds, but their detection accuracy fluctuates significantly under complex background noise.

[0004] Therefore, it is essential to develop novel algorithms that can improve detection accuracy and speed while compressing model size and reducing computational overhead in order to truly realize their practical application value. Furthermore, the system must possess strong anti-interference robustness, capable of adapting to the influence of multiple interference factors such as changes in lighting, weather fluctuations, and complex geographical environments, to ensure stable and reliable operation in various real-world scenarios. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the present invention aims to provide a sparse gated dual-domain target detection method for UAV images, which achieves synergistic optimization in three dimensions: detection accuracy, model lightweighting, and inference efficiency, effectively overcoming the limitations of existing technologies in terms of computational efficiency, small target detection, and environmental adaptability.

[0006] To achieve the objective of this invention, the following solution is adopted: A sparse-gated dual-domain target detection method for UAV images includes the following steps: Step S1: Acquire UAV image data and preprocess the UAV image data to obtain preprocessed image data; Step S2: Design a sparse gated dual-domain target detection model; The sparse gated dual-domain target detection model uses the RT-DETR model as the baseline model. The RT-DETR model includes a backbone network and an encoder; Improvements to the backbone network: The first and second stage baseline modules in the ResNet-18 backbone network are replaced with partially convolutional basic blocks; the third and fourth stage baseline modules are replaced with fast gating basic blocks; and the partially convolutional basic blocks and fast gating basic blocks are co-optimized in the spatial domain and feature domain respectively, forming a sparse gating dual-domain enhancement mechanism. Improvements to the encoder: In the cross-scale feature fusion module of the encoder, a spatially depth-separable convolution module and a cross-stage grouping convolution module are introduced. The feature map P2 output by the first-stage convolutional basic block of the improved backbone network is input into the spatially depth-separable convolution module. The feature reuse module in the encoder is replaced with a star-shaped multi-branch aggregation network module. Step S3: Using the training set in the preprocessed image data, train the sparse gated dual-domain target detection model, and obtain the optimal target detection model through iterative optimization; Step S4: Use the optimal target detection model to perform target detection on the acquired UAV images.

[0007] Furthermore, the preprocessing of the UAV image data in step S1 includes: The drone image data was uniformly cropped to 640×640 pixels; The cropped image data is divided into training set, validation set and test set according to a preset ratio.

[0008] Furthermore, the backbone network of the sparse gated dual-domain target detection model in step S2 is used to extract multi-scale feature maps from the input image, including P2, P3, P4 and P5 feature maps.

[0009] Furthermore, the spatial depth separable convolution module and the cross-stage grouped convolution module include parallel multi-dilation rate grouped convolutional layers.

[0010] Furthermore, the star-shaped multi-branch aggregation network module includes a multi-branch structure with star blocks as repeating units, and the star blocks are designed in conjunction with large kernel depth separable convolution and gating mechanism.

[0011] Furthermore, in step S3, when training the sparse gated dual-domain target detection model, the AdamW optimizer is used with a learning rate of 1×10^-4, a batch size of 150, and an input image size of 640×640.

[0012] Furthermore, in step S3, the target detection model with the best performance is retained through iterative optimization, including determining and saving the optimal target detection model based on the mAP50, mAP75 and mAP50-95 indices on the validation set and test set.

[0013] Furthermore, the drone image data acquired in step S1 includes the VisDrone2019 dataset and the HIT-UAV-V1 dataset.

[0014] Furthermore, in step S4, the optimal target detection model is used to detect targets in the UAV image, and the target bounding box and confidence score are output.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: 1. This invention achieves a synergistic improvement in detection accuracy and computational efficiency. This invention improves the original backbone network by replacing the baseline modules in the original ResNet-18 backbone network with partial convolutional basic blocks and fast gating basic blocks. In the first and second stages of the backbone network, the partial convolutional basic blocks perform convolution operations only on some input channels through a selective channel activation mechanism, while maintaining the identity mapping of the remaining channels, thereby achieving structured sparsity in the spatial-channel dimension. In the third and fourth stages of the backbone network, the fast gating basic blocks use their powerful gated linear units to dynamically and finely regulate and enhance deep semantic features, improving the nonlinear expressive power of features and effectively compensating for the information loss that may be caused by sparse operations, thus ensuring the model's representation accuracy while maintaining lightweight design. An encoder consisting of a "spatial depth separable convolutional module," a "cross-stage grouping convolutional module," and a "star-shaped multi-branch aggregation network module" is introduced, forming a systematic lightweight architecture. Experimental data shows that the final model achieves an inference speed of 80.67 FPS on the VisDrone2019 dataset, far exceeding the baseline model's 50.08 FPS, while reducing the number of model parameters by 8.2%, making it highly suitable for real-time target detection on UAV embedded platforms with limited computing resources. 2. This invention significantly enhances the detection capability for small and dense targets. The collaborative work of the "spatial depth separable convolution module," "cross-stage grouping convolution module," and "star-shaped multi-branch aggregation network module" in the encoder achieves dynamic and efficient fusion of shallow detail information and high-level semantic information. This multi-scale feature enhancement mechanism effectively solves the problem of small target feature loss caused by excessively fast downsampling in the backbone network, significantly reducing the false negative and false positive rates in complex UAV aerial photography scenarios (such as those with target occlusion, varying scales, and background interference). 3. This invention improves the model's environmental adaptability and generalization ability. The model constructed by this method enhances its robustness to different lighting, weather, and scene changes through deep optimization and adaptive fusion of multi-scale features. Test results on several challenging datasets, such as VisDrone2019 and HIT-UAV-V1, show that the proposed solution outperforms the baseline model across the board in core evaluation metrics (such as mAP50, mAP50-95, recall, etc.), demonstrating its excellent cross-dataset generalization performance and its ability to better adapt to the complex and ever-changing environments faced by UAVs in real-world applications. Attached Figure Description

[0016] Appendix Figure 1 This is a flowchart of a sparse gated dual-domain target detection method for UAV images in an embodiment of the present invention; Appendix Figure 2 This is a flowchart of image data acquisition and preprocessing in an embodiment of the present invention; Appendix Figure 3 This is a structural diagram of a sparse gated dual-domain target detection model for UAV images in an embodiment of the present invention; Appendix Figure 4 This is a structural diagram of the backbone network in an embodiment of the present invention; Appendix Figure 5 This is a structural diagram of the encoder in an embodiment of the present invention; Appendix Figure 6 This is a structural diagram of the spatial depth separable convolution module and the cross-stage grouping convolution module in an embodiment of the present invention; Appendix Figure 7 This is a diagram of a star-shaped multi-branch aggregation network module in an embodiment of the present invention; Appendix Figure 8 This is a training graph of the baseline model in an embodiment of the present invention; Appendix Figure 9 A flowchart illustrating the determination of hyperparameters, software environment, and hardware devices in embodiments of the present invention; Appendix Figure 10 This is a comparison chart of the detection performance of the embodiment of the present invention and the baseline model in complex scenes of the VisDrone2019 dataset. Detailed Implementation

[0017] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0018] As attached Figure 1 As shown, this embodiment of the invention provides a sparsely gated dual-domain target detection method for UAV images, including the following steps: Step S1: Acquire UAV image data and preprocess the UAV image data to obtain preprocessed image data.

[0019] In this embodiment, step S1 is the UAV image data acquisition and preprocessing step.

[0020] Furthermore, in step S1, VisDrone2019 was selected as the core benchmark dataset, providing the most comprehensive evaluation framework for UAV vision tasks. This dataset contains 10,209 still images and 288 video clips (a total of 261,908 frames), collected from diverse environments (including urban and rural scenes, and various weather and lighting conditions) in 14 different cities in China. Through meticulous manual annotation, over 2.6 million bounding boxes were provided, covering 10 types of targets including pedestrians and various vehicles. Key attributes such as occlusion and scene visibility were also annotated, and the acquired data underwent size cropping and data augmentation operations. (See attached image.) Figure 2 As shown, the specific steps include:

[0021] S101: Obtain the VisDrone2019 dataset.

[0022] The VisDrone2019 dataset was obtained online. This dataset is a large-scale benchmark dataset based on drone data collection. It contains 10,209 still images and 288 video clips (a total of 261,908 frames), collected from diverse environments (including urban and rural scenes, and different weather and lighting conditions) in 14 different cities in China. It primarily targets common subjects such as pedestrians, cars, and bicycles.

[0023] S102: Trim the size of the VisDrone2019 dataset.

[0024] Using the OpenCV image processing library in Python, the images were uniformly scaled to 640×640 pixels to ensure the consistency of the model input.

[0025] S103: Divide the processed images into training set, validation set and test set according to a preset ratio.

[0026] The VisDrone2019 dataset is divided into 6,471 training images, 548 validation images, and 1,610 test images, providing an authoritative benchmark for evaluating the model's generalization ability in complex real-world scenarios.

[0027] S110. To further examine the model's cross-dataset generalization ability and adaptability to special scenarios, this study introduced the HIT-UAV-V1 dataset.

[0028] The HIT-UAV-V1 focuses on infrared thermal imaging target detection for high-altitude unmanned aerial vehicles (UAVs). Its 2,898 high-resolution infrared images are all extracted from 43,470 frames of video data, covering various scenes such as schools, parking lots, roads, and playgrounds, and are labeled with target categories such as people, cars, and bicycles.

[0029] S111: Trim the size of the HIT-UAV-V1 dataset.

[0030] Using the OpenCV image processing library in Python, the images were uniformly scaled to 640×640 pixels to ensure the consistency of the model input.

[0031] S112: Divide the processed images into training set, validation set and test set according to a preset ratio.

[0032] The HIT-UAV-V1 dataset was divided into a training set (2,040 images), a validation set (287 images), and a test set (571 images).

[0033] Step S2: Design a sparse gated dual-domain target detection model.

[0034] The sparse gated dual-domain target detection model uses the RT-DETR model as the baseline model.

[0035] The RT-DETR model includes a backbone network and an encoder.

[0036] The backbone network is improved by replacing the first and second stage baseline modules in the ResNet-18 backbone network with partially convolutional basic blocks; and replacing the third and fourth stage baseline modules with fast gating basic blocks. Through the collaborative optimization of the partially convolutional basic blocks and the fast gating basic blocks in the spatial domain and feature domain respectively, a sparse gating dual-domain enhancement mechanism is formed.

[0037] Improvements to the encoder: In the cross-scale feature fusion module of the encoder, a spatial depth separable convolution module and a cross-stage grouping convolution module are introduced. The feature map P2 output by the first stage convolutional basic block of the improved backbone network is input into the spatial depth separable convolution module. The feature reuse module in the encoder is replaced with a star-shaped multi-branch aggregation network module.

[0038] The sparse gated dual-domain target detection model adopts an end-to-end query, decoding, and prediction framework.

[0039] In this embodiment, based on the RT-DETR model, a sparsely gated dual-domain target detection model for UAV images is designed through several key improvements. (See attached...) Figure 3 As shown, the specific steps include:

[0040] S201: First, the backbone network extracts multi-scale features P2, P3, P4, and P5 from the input UAV image.

[0041] S202: Next, the encoder receives the multi-scale features and introduces a spatially deep separable convolution module, a cross-stage grouping convolution module, and a star-shaped multi-branch aggregation network module to perform deep fusion and enhancement, outputting unified features rich in semantic and contextual information, and generating the initial query of the decoder accordingly.

[0042] S203: Finally, through an end-to-end query, decoding, and prediction framework, the decoder refines the initial query in multiple rounds to fully interact with image features, and the detection head finally converts the refined query into the corresponding bounding box coordinates and class confidence scores.

[0043] Furthermore, in step S201, the backbone network is responsible for extracting multi-level features from the input image. The process is illustrated in the figure below. This process progressively reduces the spatial resolution of the feature maps while increasing the number of channels to capture information from local details to global semantics. After the original convolutional layers, the backbone network generates multi-scale feature maps P2 to P5 through key stages.

[0044] As attached Figure 4 As shown, specifically, the baseline modules in the ResNet-18 backbone network are replaced with some convolutional basic blocks and fast gating basic blocks: Initial convolutional layers (layers 0-3): These four layers constitute an efficient structure. By continuously using small convolutional kernels and pooling layers, the number of channels is gradually increased while rapidly downsampling the image size, effectively capturing basic edge and texture features and preparing for subsequent deep networks.

[0045] Core feature extraction stage (layers 4-7): This is the key innovation of the backbone network. It no longer uses the original baseline modules, but introduces more advanced modules: Layers 4-5: Utilize partial convolutional basic blocks. This module performs standard convolutions on a subset of input channels using a selective channel activation mechanism, while maintaining identity mappings for the remaining channels, achieving structured sparsity in the spatial-channel dimension. This design reduces the number of model parameters and is one of the key aspects of achieving spatial domain computation compression in the "sparse gated dual-domain" architecture of this invention, ensuring efficient operation of the network even at deeper layers.

[0046] Layers 6-7 integrate the fast gating basic block. This module utilizes the dynamic modulation capability of the gated linear unit to achieve fine-grained control of information flow. As an important component of the feature domain enhancement mechanism in the "sparse gated dual-domain" architecture of this invention, it focuses on compensating for and improving feature quality at a deep network level. By enhancing the nonlinear expressive power of features, it effectively compensates for the information loss that may be caused by sparse operations, ensuring that the model maintains high accuracy while being lightweight.

[0047] Finally, the backbone network outputs four feature maps at different scales (P2, P3, P4, P5). P2 has the highest resolution and is rich in detailed information, which is beneficial for small object detection; P5 has the lowest resolution but the richest semantic information.

[0048] Furthermore, in step S202, the encoder detection head receives multi-scale features (P2, P3, P4, P5) from the backbone network. Its core task is to perform efficient intra-scale feature interaction and cross-scale feature fusion. (See attached diagram.) Figure 5 As shown:

[0049] 1. High-level feature P5 enters the feature interaction module for intra-scale feature interaction. The processed feature output is Y5. Next is a top-down feature fusion path, the purpose of which is to pass the rich semantic information of high-level features to lower layers, enhancing the expressive power of low-level features. The Y5 feature is first upsampled by the upsampling module. The upsampled feature is then concatenated with the mid-level feature P4 from the backbone network. The concatenated feature is then fed into the star-shaped multi-branch aggregation network module. The star-shaped multi-branch aggregation network module is based on a multi-branch feature reuse architecture. It can simultaneously capture features from different receptive fields through its internal multi-branch structure, thereby effectively improving the detection capability of small targets.

[0050] 2. The fusion process continues to extend to lower layers. The Y4 feature is upsampled again and then concatenated with the lower-level features P3 and P2 from the backbone network, which have been processed by the spatially depth-separable convolution module. Spatially depth-separable convolution is a type of space-to-depth convolution that achieves downsampling by rearranging pixels in the spatial dimension to the channel dimension. This method preserves information better than using stride convolution. This concatenation integrates the rich semantic information from Y4, the high-resolution details of P3, and the superior feature representation brought by spatially depth-separable convolution.

[0051] 3. Next, the fused features are sequentially passed through a cross-stage grouped convolutional module and a star-shaped multi-branch aggregation network module. The cross-stage grouped convolutional module uses grouped convolutional layers with different dilation rates to extract multi-scale features, further extracting and enhancing features, especially detailed information about small objects, while maintaining a lightweight design. The final star-shaped multi-branch aggregation network module ensures that all features are fully integrated, generating a feature map X3 rich in detail and semantic information.

[0052] 4. Simultaneously, to transmit the fine-grained localization information of the lower-level features back to the higher levels, the model adopts a bottom-up path. Feature X3 first passes through a downsampling module, and the downsampled feature is concatenated with feature Y4 from the top-level path. The concatenated feature is then fused by a star-shaped multi-branch aggregation network module. This step outputs the fused feature F4. F4 passes through the downsampling module again and is concatenated with the higher-level feature Y5. This concatenation finally integrates the detailed information from the lowest level with the semantic information from the highest level. Similarly, the concatenated feature is processed by the star-shaped multi-branch aggregation network module to generate the depth-optimized multi-scale feature F5 used by the model for object detection.

[0053] 5. At this point, the collaborative work of each module is complete, resulting in a multi-scale feature representation that combines strong semantic information with precise localization details.

[0054] Specifically, in this embodiment, a spatially depth-separable convolutional module and a cross-stage grouping convolutional module are introduced into the cross-scale feature fusion module of the encoder, specifically in conjunction with the appendix. Figure 6 Explanation. (See attached document) Figure 6 As shown, the core operations of the spatially depth-separable convolution module are sub-image segmentation and channel reassembly: the input feature map is spatially divided into non-overlapping 2×2 blocks, and the features of the four spatial locations within each block are concatenated along the channel dimension, thereby halving the spatial size while quadrupling the number of channels. This operation reduces computational resolution while preserving the original spatial information to the greatest extent, effectively preventing the loss of fine features of small targets during downsampling, and providing a high-information-density feature representation for subsequent detection tasks. The sub-image segmentation and channel reassembly operations are sparsity operations performed in the spatial dimension. Furthermore, to optimize multi-scale feature integration and enhance feature representation capabilities, a cross-stage grouped convolution module is introduced, combining the advantages of the cross-stage bias architecture in gradient flow optimization with the multi-scale feature extraction capabilities of grouped convolution.

[0055] The cross-stage grouped convolution module extracts multi-scale features by using multiple grouped convolutional layers with different dilation rates in parallel. The input features are first adjusted by a learnable bias, then simultaneously fed into three grouped convolutional layers. The output of each convolutional layer undergoes dimensionality transformation and is processed by the RPReLU activation function. Finally, the outputs of the three branches are summed and integrated through layer normalization, ultimately restoring the feature map to its original spatial dimensions. This design effectively expands the model's receptive field and aggregates contextual information at different scales without significantly increasing the number of parameters, thereby enhancing the model's feature representation capabilities.

[0056] Furthermore, in this embodiment, the feature multiplexing module in the encoder is replaced with a star-shaped multi-branch aggregation network module, specifically in conjunction with the attached diagram. Figure 7 Explanation. (See attached document) Figure 7 As shown, the MANet model based on a multi-branch feature reuse architecture replaces the original module with an innovative star-shaped block module. Specifically, the star-shaped block module first uses a 7×7 large convolutional kernel to capture long-distance spatial dependencies, and then generates two feature branches; the star operation, i.e., the element-wise multiplication fusion mechanism across subspace features, merges the two feature branches, enhancing the model's feature interaction capability. Compared with the standard convolutional bottleneck module in MANet, the star-shaped multi-branch aggregation network module uses the star-shaped block module as its core repeating unit, and its key advantages are: 1. Expanding the receptive field through large-kernel deep convolution enhances local feature extraction, a design that is more conducive to capturing fine-grained features; 2. Star operations, namely the element-wise multiplication fusion mechanism of cross-subspace features, achieve efficient and compact feature interaction; Step S3: Use the training set in the preprocessed image data to train the sparse gated dual-domain target detection model, and obtain the optimal target detection model through iterative optimization.

[0057] In this embodiment, a sparse gated dual-domain target detection model is trained using a UAV image dataset, and the optimal target detection model is retained through iterative optimization. (See attached image.) Figure 8 As shown, the specific steps include: S301: First, determine the software environment, hardware devices, and hyperparameters, including learning rate, batch size, number of iterations, and optimizer.

[0058] Batch size determination: During model training, the batch size directly affects the model's generalization ability, computational resource consumption, and final performance. Analysis of specific training logs determined the baseline model's batch size to be 150 epochs. (See attached image) Figure 9 As shown, in a training scheme with a batch size of 150, the baseline model approaches convergence at approximately 120-130 epochs. The baseline model reaches its performance peak at the 137th epoch, after which all metrics tend to stabilize.

[0059] Determining the software environment, hardware devices, and other hyperparameters: This invention was developed using Python and implemented based on the PyTorch framework. The software environment included Python 3.10.16, PyTorch 2.2.2, and CUDA 12.1. An Nvidia GeForce RTX 3090 graphics card with 24GB of VRAM was used in the experiments. During training, the input images were adjusted to a resolution of 640×640, with a batch size of 4 samples and a learning rate of 0.0001. The AdamW optimizer was used during training, with an initial learning rate of 1×10^-4, a momentum coefficient of 0.9, and a weight decay rate of 1×10^-4.

[0060] S302: Continuously train and determine and save the optimal object detection model based on metrics such as mAP50, mAP75, and mAP50-95 on the validation and test sets.

[0061] Step S4: Use the optimal target detection model to perform target detection on the acquired UAV images.

[0062] The following is in conjunction with the appendix Figure 10 The results are illustrated with data. To further test the model's effectiveness, comparative experiments were conducted on the VisDrone2019 and HIT-UAV-V1 datasets, and the results are shown in the table below.

[0063] Detection results on the VisDrone2019 dataset: Detection results on the HIT-UAV-V1 dataset: Systematic experiments on two representative UAV aerial photography datasets, VisDrone2019 and HIT-UAV-V1, demonstrate that the sparse gated dual-domain target detection model proposed in this study significantly outperforms the baseline model in both detection accuracy and inference efficiency, exhibiting excellent overall performance.

[0064] On the VisDrone2019 dataset, the sparsely gated dual-domain object detection model achieved comprehensive improvements across all key metrics. Specifically, the core metric of overall detection accuracy, mAP50, reached 0.4802, a 2.35 percentage point improvement compared to the baseline model's 0.4567; the more stringent mAP50-95 metric also improved from 0.2751 to 0.2979. Furthermore, the model's precision increased from 0.5989 to 0.6132, indicating effective control of the false detection rate; while recall significantly improved from 0.4343 to 0.4613, demonstrating a significant enhancement in the model's coverage of real targets, particularly its ability to detect small-scale and occluded targets. These data collectively demonstrate the overall strengthening of the model's target localization and classification capabilities in complex urban contexts. Moreover, this model successfully achieved its lightweight design goal while simultaneously improving performance. The number of model parameters was reduced from 19.88M to 18.25M, effectively reducing memory usage. At the same time, thanks to structural optimization, inference speed was significantly improved, with the frame rate (FPS) increasing dramatically from 50.08 frames per second to 80.67 frames per second. This characteristic of "reducing parameters and increasing speed" makes it extremely valuable for deployment on drone edge computing platforms where computing resources are scarce.

[0065] Cross-dataset validation on the HIT-UAV-V1 dataset further confirms the powerful generalization ability of the sparsely gated dual-domain object detection model. While maintaining high precision (0.8748), the model outperforms the baseline in recall (0.7427) and the overall metric F1-Score (0.7930), indicating more balanced and reliable performance across different scene distributions. mAP50 improves to 0.7659, and mAP50-95 improves to 0.5048, demonstrating that the algorithm can effectively adapt to different data characteristics and robustly handle typical challenges in UAV applications such as illumination variations, scale differences, and partial occlusion, showing promising practical application prospects.

[0066] To demonstrate the effects achieved by the invention, in conjunction with the appendix Figure 10 As shown, the comparison of the detection performance of the sparse gated dual-domain object detection model and the baseline model in complex scenes on the VisDrone2019 dataset is presented: In terms of preventing missed detections in dense scenarios, the sparse gating dual-domain target detection model demonstrates significant advantages. (See attached image) Figure 10 As shown in (a), in urban road environments with severe tree obstruction, the baseline model failed to identify vehicle targets partially obscured by vegetation, while the sparse gating dual-domain target detection model successfully captured these difficult samples, achieving accurate localization and classification.

[0067] The sparsely gated dual-domain target detection model demonstrates superior adaptability in small target detection and multi-scale target collaborative processing. Quantitative evaluation shows that this model detects a significantly larger number of small-pixel targets than the baseline model. (See attached image) Figure 10 As shown in (b), the sparse gated dual-domain target detection model accurately identifies the near large targets, tricycles and motor vehicles, which were misclassified by the baseline model, and successfully detects the distant small targets, motor vehicles, which were not detected by the baseline model.

[0068] The sparse gated dual-domain target detection model demonstrates strong anti-interference performance in distinguishing between overlapping targets and complex background interference. (See attached image) Figure 10 As shown in (c), in the high-density vehicle group detection task, facing the complex situation of severe occlusion and target overlap, the sparse gating dual-domain target detection model can still stably detect most mutually occluded vehicle targets, while the baseline model shows obvious missed detections under the same conditions.

[0069] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A sparse-gated dual-domain target detection method for UAV images, characterized in that, Includes the following steps: Step S1: Acquire UAV image data and preprocess the UAV image data to obtain preprocessed image data; Step S2: Design a sparse gated dual-domain target detection model; The sparse gated dual-domain target detection model uses the RT-DETR model as the baseline model. The RT-DETR model includes a backbone network and an encoder; Improvements to the backbone network: The first and second stage baseline modules in the ResNet-18 backbone network are replaced with partially convolutional basic blocks; the third and fourth stage baseline modules are replaced with fast gating basic blocks; and the partially convolutional basic blocks and fast gating basic blocks are co-optimized in the spatial domain and feature domain respectively, forming a sparse gating dual-domain enhancement mechanism. Improvements to the encoder: In the cross-scale feature fusion module of the encoder, a spatially depth-separable convolution module and a cross-stage grouping convolution module are introduced. The feature map P2 output by the first-stage convolutional basic block of the improved backbone network is input into the spatially depth-separable convolution module. The feature reuse module in the encoder is replaced with a star-shaped multi-branch aggregation network module. The sparse gated dual-domain target detection model adopts an end-to-end query, decoding, and prediction framework; Step S3: Using the training set in the preprocessed image data, train the sparse gated dual-domain target detection model, and obtain the optimal target detection model through iterative optimization; Step S4: Use the optimal target detection model to perform target detection on the acquired UAV images.

2. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, The preprocessing of the UAV image data in step S1 includes: The drone image data was uniformly cropped to 640×640 pixels; The cropped image data is divided into training set, validation set and test set according to a preset ratio.

3. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, The backbone network of the sparse gated dual-domain target detection model in step S2 is used to extract multi-scale feature maps from the input image, including P2, P3, P4 and P5 feature maps.

4. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, The spatial depth separable convolutional module and the cross-stage grouped convolutional module include parallel multi-dilation rate grouped convolutional layers.

5. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, The star-shaped multi-branch aggregation network module includes a multi-branch structure with star blocks as repeating units. The star blocks are designed in conjunction with large kernel depth separable convolution and gating mechanism.

6. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, In step S3, when training the sparse gated dual-domain target detection model, the AdamW optimizer is used with a learning rate of 1×10^-4, a batch size of 150, and an input image size of 640×640.

7. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, In step S3, the target detection model with the best performance is retained through iterative optimization, including determining and saving the optimal target detection model based on the mAP50, mAP75 and mAP50-95 indices on the validation set and test set.

8. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, The UAV image data acquired in step S1 includes the VisDrone2019 dataset and the HIT-UAV-V1 dataset.

9. The sparse gated dual-domain target detection method for UAV images according to claim 1, characterized in that, In step S4, the optimal target detection model is used to detect targets in the UAV image, and the target bounding box and confidence score are output.

Images