A target detection method, device and equipment for unmanned aerial vehicle visible light images
By using a feature decoupling module and a differential-guided feature aggregation module, combined with dynamic matchability perception loss, the problems of spatial-semantic entanglement and improper multi-scale feature fusion in UAV target detection are solved, achieving lightweight and efficient small target detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XINYANG NORMAL UNIVERSITY
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-12
Smart Images
Figure CN122200450A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target detection technology, and in particular to a target detection method, apparatus, equipment and medium for visible light images of unmanned aerial vehicles. Background Technology
[0002] The rapid proliferation of unmanned aerial vehicles (UAVs) in civilian, industrial, and public safety fields has driven the growing demand for intelligent aerial sensing systems. UAV-based target detection (UAV-OD) is a core technology for achieving autonomous scene understanding, supporting applications such as infrastructure inspection, traffic monitoring, emergency response, and environmental surveillance. Compared with traditional target detection, UAV-OD faces unique challenges: images acquired from a low-altitude, top-down perspective typically contain extremely small and densely distributed targets, severe background clutter, and significant scale and viewpoint variations. Furthermore, airborne computing and storage resources are often limited, making the deployment of high-capacity detectors impractical. These characteristics collectively make UAV-OD a challenging problem that requires simultaneously achieving high accuracy and real-time efficiency for small targets.
[0003] Current UAV target detection methods mainly rely on one-stage and two-stage CNN-based frameworks. One-stage CNN detectors emphasize dense prediction and real-time inference, but often perform poorly in locating small targets, and cross-scale features are easily diluted during fusion. Two-stage detectors improve localization accuracy through candidate regions and multi-stage refinement, but at the cost of higher computational cost and greater latency.
[0004] To address the issues in one-stage and two-stage CNN-based frameworks, the DETR-based end-to-end detector utilizes an attention mechanism to capture global dependencies, avoiding anchor box design and non-maximum suppression, thus simplifying the detection process. However, the DETR detector's self-attention mechanism simultaneously models spatial structure and semantic context in the global feature space. Due to the extremely small pixel proportion of small targets, conflicts often arise between spatial structure and semantic context, causing subtle edge information to be submerged by semantic information. This makes it difficult to accurately extract detailed edge features and semantic features of small targets in parallel. Furthermore, when aggregating target features after extraction, the DETR detector typically uses self-attention to fuse features from different layers with uniform or static weights. Low-level features are rich in details crucial to small targets but contain a lot of background noise, while high-level features are semantically abstract but have coarse spatial information. Uniform fusion causes noise to propagate along with details, and semantic information is difficult to accurately guide detail enhancement. As a result, the fused features cannot accurately represent the global features of small targets, ultimately making it difficult to detect small targets with high accuracy. Summary of the Invention
[0005] This invention provides a target detection method, apparatus, and device for visible light images of unmanned aerial vehicles (UAVs), which can solve the problems existing in the prior art.
[0006] This invention provides a target detection method for visible light images from unmanned aerial vehicles (UAVs), comprising the following steps: Acquire visible light images captured by a drone; A visible light image is input into a pre-trained target detection model, which includes a backbone network, a feature decoupling module, and a feature aggregation module. Multi-scale feature maps of visible light images are extracted through a backbone network. A spatial guidance branch of the feature decoupling module performs differential Gaussian operations on the feature maps at each scale to obtain high-frequency residual features representing edge texture. A spatial attention mechanism is used to generate a spatial attention weight map of the high-frequency residual features. The spatial attention weight map is then multiplied element-wise with the feature map at the corresponding scale to obtain spatial edge texture features. A context-aware branch of the feature decoupling module uses large-kernel convolution to aggregate contextual information within the local regions expanded by the feature maps at each scale to obtain region aggregation features. Global average pooling is used to obtain the channel semantic statistical vectors of the feature maps at the corresponding scale. The region aggregation features and channel semantic statistical vectors are then added element-wise and fused to obtain a semantic attention weight map. Finally, the semantic attention weight map is multiplied element-wise with the feature map at the corresponding scale to obtain semantic context features. Then, the spatial edge texture features corresponding to each scale feature map are added and fused with the semantic context features element by element to obtain the decoupling enhancement features corresponding to each scale; The feature aggregation module processes adjacent scale pairs step by step along the feature pyramid, extracts the feature difference between the decoupled enhancement features of each adjacent scale pair as a guiding signal, generates channel modulation weights through a lightweight adaptive controller based on the guiding signal, and uses the channel modulation weights to perform weighted fusion of the decoupled enhancement features of the scale pair to obtain the intermediate aggregated features of the scale pair. The intermediate aggregated features are used as the new current scale and continue to perform differential guided aggregation with the decoupled enhancement features of the next adjacent scale to obtain the final aggregated features. Targets in visible light images are detected based on the final aggregated features.
[0007] Preferably, the acquisition of the spatial edge texture features includes: The multi-scale feature map of the visible light image is as follows: ,make and For channel-by-channel Gaussian kernels applied independently, and whose standard deviations satisfy... The high-frequency response of capturing edges and textures is represented as: Where: represents the convolution operation; Two channel-wise Gaussian kernels and Using a single Gaussian kernel By approximation, the high-frequency residual calculation is expressed as: After obtaining the high-frequency residual, use Convolution for local structural refinement and noise suppression is represented as follows: After local structural refinement and noise suppression, pointwise linear dimensionality reduction-recovery of the PWLR bottleneck is used to adaptively weight spatial location and channels to obtain a spatial attention weight map. Let the dimensionality reduction ratio be... , ,but: in: , Indicates learnable Convolution kernel; Represents the Sigmoid function; Applying the spatial attention weight map to the input features yields spatial edge texture features, represented as: in: This indicates element-wise multiplication.
[0008] Preferably, the acquisition of the semantic context features includes: Using large-kernel depthwise convolutions, contextual information is aggregated within local regions expanded at each scale feature map to obtain region aggregated features, represented as: in: For learnable Convolution kernel; Global channel-level semantic priors are obtained through global average pooling, resulting in channel semantic statistical vectors, represented as follows: in: This represents the average activation of each channel in the spatial domain; The region aggregation features and channel semantic statistical vectors are transformed using the PWLR bottleneck and expressed as follows: The transformed features are fused element-wise to obtain the semantic attention weight map, represented as: Applying the semantic attention weight map to the original features yields semantic context features, represented as follows: in: This indicates element-wise multiplication.
[0009] Preferably, obtaining the intermediate aggregation feature includes: make , The decoupling enhancement features at two scales are represented by projecting both onto a shared low-dimensional subspace using low-rank convolution, as follows: in: ; Differential Feature Map The differential feature map is used as a guiding signal for feature aggregation: In top-down propagation, i.e., higher levels to lower level During the dissemination, This indicates the missing fine-grained details in high-level features, guiding the module to selectively inject low-level spatial information; In bottom-up aggregation, i.e., the lower layer To the upper levels During aggregation, Capture supplementary semantic abstractions in low-level features to enable selective reinforcement of high-level semantic cues; The channel-level global descriptor is obtained through explicit spatial averaging, and is represented as follows: in: ; The global descriptor is obtained by concatenating the three elements. The global descriptor is then input into the lightweight two-layer controller to generate channel modulation weights, represented as: in: Corresponding to and Channel weights; Modulating and fusing the features is represented as: Local refinement is achieved using convolution and residual connections are introduced to obtain intermediate aggregated features, represented as follows: in: These are learnable residual scaling parameters.
[0010] Preferably, during training, the target detection model uses a Dynamic Matchability-Aware Loss (DMA-Loss), which includes: The first The loss for each query is defined as: in: The predicted prospect probability; q i The focus index represents the positioning quality measured by the intersection-union ratio (IU). Based on the dynamic determination of positioning quality, it is expressed as: in: Control the sensitivity to positioning errors; Set the minimum focus intensity; Loss through location quality q i The relevant gradient modulation for two-dimensional classification is represented as follows: when hour, Smaller The losses were mainly caused by Dominant, used to enhance foreground supervision of reliable positive samples; when hour, Increase inhibition And enhance the background-related terms to weaken the supervision strength of unreliable positive samples, in order to reduce the noise gradient caused by low-quality matching; when hour, Decrease Amplification A stronger penalty is applied to high-confidence false positives that are close to the real target, which is used to distinguish the foreground from the background; when hour, Increase inhibition This is used to focus training on more informative samples.
[0011] Preferably, the target detection model adopts an encoder-decoder architecture, wherein the backbone network adopts the HGNetv2 backbone network.
[0012] This invention also provides a target detection device for visible light images of unmanned aerial vehicles, comprising: Image module, used to acquire visible light images captured by the drone; The detection module is used to input visible light images into a pre-trained target detection model, which includes a backbone network, a feature decoupling module, and a feature aggregation module. Multi-scale feature maps of visible light images are extracted through a backbone network. A spatial guidance branch of the feature decoupling module performs differential Gaussian operations on the feature maps at each scale to obtain high-frequency residual features representing edge texture. A spatial attention mechanism is used to generate a spatial attention weight map of the high-frequency residual features. The spatial attention weight map is then multiplied element-wise with the feature map at the corresponding scale to obtain spatial edge texture features. A context-aware branch of the feature decoupling module uses large-kernel depthwise convolution to aggregate contextual information within the local regions expanded by the feature maps at each scale to obtain region aggregation features. Global average pooling is used to obtain the channel semantic statistical vectors of the feature maps at the corresponding scale. The region aggregation features and channel semantic statistical vectors are then added element-wise and fused to obtain a semantic attention weight map. Finally, the semantic attention weight map is multiplied element-wise with the feature map at the corresponding scale to obtain semantic context features. Then, the spatial edge texture features corresponding to each scale feature map are added and fused with the semantic context features element by element to obtain the decoupling enhancement features corresponding to each scale; The feature aggregation module processes adjacent scale pairs step by step along the feature pyramid, extracts the feature difference between the decoupled enhancement features of each adjacent scale pair as a guiding signal, generates channel modulation weights through a lightweight adaptive controller based on the guiding signal, and uses the channel modulation weights to perform weighted fusion of the decoupled enhancement features of the scale pair to obtain the intermediate aggregated features of the scale pair. The intermediate aggregated features are used as the new current scale and continue to perform differential guided aggregation with the decoupled enhancement features of the next adjacent scale to obtain the final aggregated features. Targets in visible light images are detected based on the final aggregated features.
[0013] This invention also provides an electronic device, including a memory and a processor; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of a target detection method for visible light images of unmanned aerial vehicles as described above.
[0014] This invention also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of a target detection method for visible light images of unmanned aerial vehicles as described above.
[0015] This invention provides a target detection method, apparatus, and device for visible light images from unmanned aerial vehicles (UAVs). Compared with existing technologies, its advantages are as follows: Inspired by the dual-path perception mechanism of the human visual system, this invention constructs an explicit feature decoupling module in the feature extraction stage. This module captures long-range semantic dependencies by establishing a context-aware branch corresponding to the ventral pathway and simultaneously establishing a spatial guidance branch corresponding to the dorsal pathway to preserve fine structural details. This eliminates representational ambiguities between target identification and spatial localization within independent subspaces, thus avoiding conflicts between spatial structure and semantic context. Subsequently, during feature aggregation, since the information transfer between multi-scale feature layers is essentially a directional and purposeful compensation process, a differentially guided feature aggregation mode is constructed. This mode quantifies which information is lost or abstracted from one layer to another, allowing for targeted infusion or selective pumping of information. It abandons the brute-force calculation of all position pairs by self-attention and dynamically adjusts the information flow to bypass or suppress unreliable feature regions for accurate aggregation, obtaining global features of small targets for final target detection.
[0016] Furthermore, during the model training phase, this invention constructs a Dynamic Matchability-Aware Loss (DMA-Loss) and combines the alignment difference between classification confidence and localization quality to perform differentiated gradient gain adjustment on samples in different learning behavior states, thereby eliminating noise gradients caused by inconsistent matching and achieving focused optimization of the training process. Attached Figure Description
[0017] Figure 1 This is a diagram comparing the performance and computational complexity of representative real-time and end-to-end target detectors in small target detection at present. Figure 2 This is a schematic diagram of the overall structure of the Fast-DETR method for target detection in visible light images from unmanned aerial vehicles, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the overall structure of an SCDM (Supervisory Control Method) for target detection in visible light images from unmanned aerial vehicles, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the overall structure of a DGFA (Distributed Gaussian Graphical Application) method for target detection in visible light images from unmanned aerial vehicles, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the DMA-Loss mechanism for a target detection method for visible light images of UAVs provided in an embodiment of the present invention; Figure 6 Qualitative detection results and corresponding response heatmap of the STATE-AIR dataset for a target detection method for visible light images of unmanned aerial vehicles provided in this embodiment of the invention; Figure 7 Qualitative detection results and corresponding response thermal diagram of the AU-AIR dataset for a target detection method for visible light images of unmanned aerial vehicles provided in this embodiment of the invention; Figure 8 Qualitative detection results and corresponding response heatmap of the VISDRONE dataset for a target detection method for visible light images of unmanned aerial vehicles provided in this embodiment of the invention; Figure 9 This is a schematic diagram illustrating the robustness experiment of a target detection method for visible light images from unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention. Detailed Implementation
[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0019] In recent years, DETR-style end-to-end detectors have been increasingly used in UAV detection tasks due to their powerful global reasoning capabilities and simpler detection processes. To further improve DETR performance, researchers have proposed several strategies, including: lightweight real-time DETR variants with resource-aware feature fusion for resource-constrained scenarios; multi-scale cross-layer pyramid designs that preserve small target cues; and training strategies for stable bipartite graph matching when small target supervision noise is high. Despite these advances, the application of existing DETR methods in low-altitude UAV images still faces significant challenges, especially in the efficiency-accuracy tradeoff in small target scenarios.
[0020] like Figure 1 As shown, although representative DETR-style detectors are competitive in overall accuracy ( Figure 1 (a)), but its small target indicator AP S A significant decrease was observed ( Figure 1 (b) Current approaches to improving small target detection performance typically introduce high computational complexity, which is reflected in... Figure 1 (c) shows a significant increase in the number of parameters and GFLOPs, while inference efficiency also decreases, such as Figure 1 As shown in (d).
[0021] These empirical observations can be attributed to three key technical bottlenecks in DETR-style UAV detectors. First, there is a spatial-semantic entanglement problem in the representation stage: the backbone network often favors high-level semantic abstraction at the expense of fine-grained spatial structure. This entanglement is often amplified by the global receptive field of self-attention, thus blurring the clear geometric boundaries required for small target localization, especially when the foreground target occupies only a very small proportion of the image. Second, the multi-scale aggregation stage is often affected by undirected scale interference: existing mechanisms usually lack sufficient scale perception and are unable to cope with the scale changes caused by drastic changes in flight altitude and viewpoint in UAV target detection. Without utilizing feature difference as a guiding signal, traditional indiscriminate propagation triggers a double negative effect: 1) the dominant high-level semantics suppresses and drowns out the fragile small target responses in the shallow layers; 2) messy background noise is propagated upwards indiscriminately, thus contaminating deep semantic features. Third, there is supervised ambiguity in the optimization stage: this problem stems from set-based matching and static allocation strategies, which ignore the dynamic changes in sample matchability during training. Specifically, poorly positioned foreground matching with low intersection-union ratios can generate noisy gradients that conflict with well-positioned predictions, leading to instability in the optimization process and weakening the recall capability for small targets.
[0022] To address the aforementioned issues, this invention proposes a lightweight Fast-DETR framework for airborne detection of small targets in low-altitude UAV images, balancing accuracy and efficiency. Fast-DETR progressively improves detection capabilities through three collaborative stages: representation, aggregation, and optimization. In the representation stage, this invention designs a Spatial-Context Feature Decoupling Module (SCDM) to decouple fine-grained spatial details from high-level semantic context, preserving accurate target boundaries while suppressing background interference. In the aggregation stage, this invention proposes a Differential Guided Feature Aggregation Module (DGFA) that selectively fuses multi-scale features based on inter-layer differences. This design efficiently propagates global semantic context and preserves key spatial structures without introducing the high computational cost of full self-attention. In the optimization stage, this invention proposes Dynamic Matchability-Aware Loss (DMA-Loss), adaptively adjusting the supervision strength based on localization quality: weakening the gradient of uncertain positive samples and emphasizing negative samples with higher information content, thereby stabilizing training and improving the detection performance for small and dense targets. Specifically:
[0023] I. Overall Architecture.
[0024] Fast-DETR is designed as a multi-level enhancement and alignment framework to mitigate the inherent perceptual interference of DETR-type detectors in complex aerial and remote sensing scenarios. This interference occurs across three tightly coupled stages: (i) at the representation level, where spatial details and semantic context are intertwined; (ii) at the feature aggregation level, where multi-scale information is encoded and propagated; and (iii) at the optimization level, where static matching strategies cannot reflect the differences in matchability among samples.
[0025] like Figure 2 As shown, Fast-DETR employs a standard encoder-decoder detection paradigm and is based on a convolutional backbone network. This invention chooses HGNetv2 as the backbone network due to its lightweight structure and strong feature extraction capabilities. Given an input image, the backbone network outputs a set of multi-scale feature maps. First, it undergoes SCDM processing deployed across multiple stages. Each SCDM explicitly decouples spatial detail cues from semantic context information, and then performs controlled intra-stage fusion on this basis to obtain an enhanced feature representation with less perceptual interference. These SCDMs together generate a set of "decoupled sensing" multi-scale feature layers, which serve as the input to the encoder.
[0026] Subsequently, the obtained features are fed into DGFA, which forms the core of the encoder. DGFA does not rely on self-attention, but performs orientation-aware cross-scale encoding by modeling the differential evolution between features in adjacent stages. With the help of bidirectional propagation, DGFA aggregates global semantic context while preserving fine-grained spatial structure, thereby achieving efficient multi-scale representation learning with a lightweight computational budget.
[0027] Finally, the encoded features are decoded by a DETR-style decoder to produce target predictions and supervised by the proposed DMA-Loss. DMA-Loss dynamically reallocates classification gradients based on the matchability of the samples, aligning the optimization process with the localization-aware representation, thus forming an end-to-end framework from representation learning to detection.
[0028] II. Spatial-Context Feature Decoupling Module (SCDM).
[0029] 1) Motivation and design principles.
[0030] At the representation level, DETR-based detectors need to simultaneously encode fine-grained spatial structure and high-level semantic context in shared feature embeddings. However, these two types of perceptual cues are inherently heterogeneous: spatial structure is mainly dominated by local geometric changes and high-frequency details, while semantic context emphasizes regional consistency and long-range dependencies. When they are entangled in the same representation space, their conflicting inductive biases often interfere with each other, resulting in blurred target boundaries and unstable attention responses, especially in small and dense target scenes.
[0031] From a perceptual mechanism perspective, the human visual system typically mitigates ambiguity through two functionally independent pathways: the ventral pathway for target identification and the dorsal pathway for spatial localization. These two pathways operate largely decoupled before fusion, thus achieving robust semantic recognition while preserving fine-grained spatial structure. Inspired by this biological principle, this invention proposes SCDM, which explicitly decouples feature representation into two functionally specialized sub-branches: (i) a spatially guided branch (SGB) for preserving structural details; and (ii) a context-aware branch (CAB) for capturing long-range semantic dependencies. The two branches are independently optimized in their respective subspaces to avoid mutual interference, and are ultimately integrated through residuals.
[0032] Given an input multi-scale feature map SCDM calculation is expressed as: (1) in: Indicates learnable weights Spatial convolution, This indicates that Batch Normalization is followed by non-linear activation; residual connections are used to preserve the learned spatial and semantic information in the backbone network. The overall framework is as follows: Figure 3 As shown.
[0033] 2) Spatial Bootstrap Branch (SGB).
[0034] This invention proposes SGB, which explicitly utilizes fine-grained structural cues in the input feature map to generate spatial attention weights, thereby emphasizing edges, boundaries, and local textures, such as... Figure 3 As shown in (a).
[0035] To isolate a more information-rich spatial structure, the SGB could originally perform differential Gaussian operations; let and For channel-by-channel Gaussian kernels applied independently, and whose standard deviations satisfy... The high-frequency response used to capture edges and textures is defined as follows:
[0036] (2) This operation emphasizes local changes (including boundaries and fine structures) while suppressing slowly changing background information; to reduce computational overhead and model complexity, we use a single Gaussian kernel. An approximation is made, and the high-frequency residual calculation is expressed as: (3) Then use Convolution for local structural refinement and noise suppression is represented as follows: (4) Despite the simplification, this single-core solution still effectively highlights spatial details while maintaining a lightweight module.
[0037] Next, SGB uses a pointwise linear dimensionality reduction-recovery (PWLR) bottleneck to adaptively weight spatial location and channels. Let the dimensionality reduction ratio be... , ,but:
[0038] (5) in , For learnable Convolutional kernels are used for dimensionality reduction and dimensionality recovery, respectively. This is the Sigmoid function.
[0039] Finally, applying the spatial attention map to the input features yields the spatially enhanced output representation as follows: (6) in: This indicates element-wise multiplication.
[0040] 3) Context-aware branch (CAB).
[0041] CAB aims to model long-range semantic context in a lightweight manner, avoiding the secondary complexity and sensitivity to background (e.g., ...) of self-attention mechanisms commonly used in DETR-based detectors. Figure 3 (b) CAB does not explicitly compute pairwise interactions, but instead models semantic context as a "semantic reweighting" problem: selectively enhancing context-consistent responses based on complementary semantic cues; specifically, CAB assumes that semantic context comes from two complementary sources: (i) extended spatial interactions, used to encode regional semantic consistency beyond local neighborhoods; (ii) global channel-level semantic priors, reflecting the overall presence strength of semantic concepts at the scene level; therefore, CAB extracts semantic information through two parallel paths.
[0042] The first path uses large-kernel convolution to aggregate and expand the spatial context, represented as: (7) in: For learnable Convolution kernel; to balance context coverage and computational efficiency, set It provides a sufficiently large effective receptive field for regional semantic aggregation without introducing excessive overhead.
[0043] In parallel, global channel-level semantic priors are obtained through spatial averaging, as follows: (8) in: This represents the average activation of each channel in the spatial domain. This global descriptor captures scene-level contextual statistics, providing a compact semantic prior to the importance of different channels.
[0044] Similar to SGB, the outputs of the two branches mentioned above are transformed through the PWLR bottleneck, as follows: (9) (10) The transformed features encode semantic importance from a complementary perspective, and a unified semantic attention map is obtained by element-wise addition and fusion, as shown below: (11) Finally, the semantic attention map is applied to the original features to obtain the context recalibration output, represented as: (12) This process selectively enhances context-consistent semantic responses while suppressing context- or semantically irrelevant activations.
[0045] III. Differential Guided Feature Aggregation (DGFA).
[0046] 1) Motivation and design principles.
[0047] At the feature aggregation level, Transformer-based encoders typically rely on self-attention to capture global contextual dependencies, but their quadratic computational complexity... This severely limits deployment on resource-constrained drone platforms. Unlike explicit modeling of all pairwise interactions, it was observed that contextual propagation in multi-scale feature hierarchies naturally follows a "directed, gradual evolution" process. Specifically, top-down propagation corresponds to a "semantic-to-detail" compensation process, where high-level semantic representations guide the recovery of missing fine-grained spatial structures. Conversely, bottom-up aggregation corresponds to an "detail-to-semantic" abstraction process, where local spatial cues are gradually refined into compact and more discriminative semantic representations. Crucially, this cross-layer semantic evolution can be explicitly characterized by feature differences between adjacent stages: these differences encode the "change and direction" of information during propagation.
[0048] Based on this observation, this invention proposes DGFA and uses it to construct an encoder without self-attention; the core idea of DGFA is to perform cross-layer feature difference. As explicit guiding signals, they are used to dynamically indicate the direction and magnitude of information propagation; these signals drive a lightweight adaptive controller to generate fusion weights, thereby enabling efficient, targeted context modeling without global attention computation.
[0049] 2) Module structure and formalization.
[0050] DGFA is a general-purpose feature aggregation unit used to model directional context propagation between adjacent feature stages through differential guidance, such as... Figure 4 As shown.
[0051] make , The enhanced features of the two inputs (whose perceptual interference has been reduced) are represented as follows: To suppress channel redundancy and align the representation space, both are first projected into a shared low-dimensional subspace through low-rank convolution, as shown below: (13) in: Subsequently, a difference feature map is constructed. This difference explicitly characterizes the differences in representation across stages and serves as a guiding signal for DGFA: propagation from top to bottom (high level) to lower level )middle, This indicates the missing fine-grained details in high-level features, guiding the module to selectively inject low-level spatial information.
[0052] Converging from bottom to top (lower level) To the upper levels )middle, Capturing supplementary semantic abstractions in low-level features allows for selective enhancement of high-level semantic cues.
[0053] The channel-level global descriptor is obtained through explicit spatial averaging, and is represented as follows: (14) (15) (16) in: The three components are then concatenated to obtain the global descriptor. And input into the lightweight two-layer controller to generate channel modulation weights, expressed as: (17) (18) in: Corresponding to and The channel weights; the features are modulated and fused, represented as: (19) Finally, convolution is used for local thinning and residual connections are introduced, as follows: (20) in: These are learnable residual scaling parameters. Output Adaptive fusion is achieved under the guidance of differential signals, enabling the target features to obtain complementary information from the source features in both top-down and bottom-up propagation.
[0054] IV. Dynamic Matchability Perceived Loss (DMA-Loss).
[0055] 1) Design motivation and principles.
[0056] Small target detection in low-altitude drone images remains challenging due to weak visual cues that are easily confused with the surrounding background; during training, matched predictions have varying localization quality, and classification confidence is not always consistent with spatial alignment; applying uniform classification supervision to all samples may introduce noisy gradients and hinder effective optimization.
[0057] Considering the actual classification labels Positioning quality as measured by IoU The study observed that the samples exhibited different learning behaviors, indicating that a single, uniform objective was insufficient to adequately address these issues; for example... Figure 5 As shown, four typical scenarios will occur during training: (1) and High (top right): The sample corresponds to an accurate foreground prediction. The classification confidence is consistent with the positioning accuracy. It is a reliable learning signal. It should be subjected to positive supervision to maintain a stable gradient and enhance the foreground discrimination features.
[0058] (2) and Low (top left): Although marked as foreground, low localization quality indicates uncertain or noisy matching, often caused by weak target cues or background interference. In this case, strong positive supervision may propagate misleading gradients and destroy training stability. Therefore, the strength of positive supervision should be reduced, while complementary background constraints are needed to help it correct from incorrect foreground predictions.
[0059] (3) and High (bottom right): The sample is a background prediction but it overlaps significantly with the real target space, making it very easy to be confused with small targets; if the emphasis is insufficient, it is easy to produce persistent false positives and cause the foreground-background boundary to become blurred. Therefore, strong negative supervision should be applied to suppress false activation and enhance separability.
[0060] (4) and Low (bottom left): Background samples that have almost no overlap with any target are usually easy to classify, but have limited information for boundary refinement; overemphasizing them will allow them to dominate gradient updates without providing effective discrimination clues, so they should be appropriately deweighted to achieve more efficient and focused optimization.
[0061] 2) Definition and discussion of loss.
[0062] follow Figure 5 The classification design principles shown, the first The DMA loss for a query is defined as: (twenty one) in: Focusing on the predicted prospect probability, the index is used. Based on the dynamic determination of positioning quality, it is expressed as: (twenty two) in: Controlling sensitivity to positioning errors, Set a minimum focus intensity. DMA loss is achieved through IoU-related gradient modulation. Figure 5 The two-dimensional classification principle.
[0063] (1) When hour, Smaller The losses were mainly caused by Leading the way, strengthen prospect monitoring of reliable positive samples.
[0064] (2) When hour, Increase inhibition It also enhances background-related terms, thereby weakening the supervision strength of unreliable positive samples and reducing the noise gradient caused by low-quality matching.
[0065] (3) When hour, Decrease Amplification Strengthen the penalty for high-confidence false positives that are close to the real target, thereby improving foreground-background separability.
[0066] (4) When hour, Increase inhibition This avoids negative samples dominating optimization and allows training to focus on more informative samples.
[0067] This adaptive gradient assignment mechanism makes classification learning more stable and effective under severe localization uncertainty.
[0068] Specific experiment: A. Dataset and evaluation metric settings.
[0069] 1) Datasets: This invention evaluates Fast-DETR on three representative low-altitude UAV benchmark datasets: State-Air, VisDrone, and AU-AIR. State-Air is a self-collected dataset containing 2,864 aerial images, of which 2,246 were taken in clear weather and 616 in snowy conditions. AU-AIR is a multimodal aerial dataset containing 32,823 RGB images collected during real flight missions, covering eight common target classes and dynamic perspectives, making it particularly suitable for evaluating model robustness in real UAV monitoring scenarios. VisDrone contains 6,471 training images, 548 validation images, and 3,190 test images, all collected by UAVs at different locations and flight altitudes.
[0070] 2) Experimental details: Fast-DETR is implemented based on the PyTorch framework and optimized using AdamW. The initial learning rate is... The weight decays to The batch size is set to 8, and all models are trained for 200 epochs. Each dataset is divided into training and validation sets in an 8:2 ratio.
[0071] 3) Evaluation Indicators: Following existing work, report standard COCO-style indicators, including AP (IoU=0.5:0.95), AP 50 (IoU=0.5) and AP 75 (IoU=0.75), and the scale-aware index AP S AP M APL Each corresponds to a smaller target area ,lie in interval, and greater than The target of the pixel.
[0072] B. Comparison with the SOTA method.
[0073] 1) Quantitative comparison: To comprehensively evaluate the detection performance of Fast-DETR, this invention is compared with 12 representative detection methods, including YOLOv11, YOLOv12, YOLOv13, FBRT-YOLO, EDNet, Faster R-CNN, Dynamic R-CNN, RT-DETR, DINO, UAV-DETR, D-FINE, and DEIM; based on this, detailed quantitative results on three UAV datasets are summarized in Table 1.
[0074] Table 1. Quantitative comparison on the State-Air, AU-Air, and VisDrone datasets. As shown in Table 1, Fast-DETR achieves the best overall performance on the State-Air dataset and consistently outperforms competing methods on most evaluation metrics; in particular, Fast-DETR achieves an AP of 52.7%, a 1.1 percentage point improvement over the strongest baseline D-FINE, while also demonstrating strong performance in AP. 75 (+1.6) and AP S Gains were also achieved at (+0.9); these improvements indicate that Fast-DETR is better at maintaining discriminative representations of small targets in scenarios with strong scale variations and background dominance, which are common in large-scale aerial images; for large targets, Fast-DETR's AP... L The accuracy was 79.0%, slightly lower than RT-DETR. This indicates that the end-to-end DETR-style detector remains competitive on dominant large targets, while Fast-DETR achieves more balanced performance across different scales.
[0075] The AU-AIR dataset is characterized by extremely small targets and sparse foreground instances, posing a greater challenge to reliable detection. On this dataset, Fast-DETR demonstrates a significant advantage over all comparable methods, achieving the highest AP (17.5%) and AP2. 50 (39.6%) and AP 75 (12.5%); Notably, Fast-DETR achieved the best AP on small targets. S(2.0%); Compared to its closest competitor DEIM, Fast-DETR achieved consistent improvements across all metrics, demonstrating stronger robustness to weak target cues and foreground-background blurred regions. Despite its AP L Slightly lower than YOLOv13 (0.2% lower), but the difference is negligible, indicating that the proposed framework enhances small / medium-scale detection without sacrificing performance on large targets.
[0076] VisDrone is one of the most challenging UAV benchmarks, characterized by dense target distribution, frequent occlusion, and extremely cluttered backgrounds. As shown in Table 1, Fast-DETR achieved the highest AP (26.1%), and... 50 AP 75 AP S With AP M It consistently outperforms all competing methods; in particular, its small-objective performance improvement (APS=18.0%) highlights the effectiveness of Fast-DETR in mitigating feature interference in dense and complex scenes. Although its AP... L Slightly lower than D-FINE, but overall results indicate that Fast-DETR is more inclined to balanced detection across scales rather than favoring large targets.
[0077] 2) Visual comparison: This invention provides a visual evaluation of the detection performance of Fast-DETR and selects YOLOv12, YOLOv13, D-FINE and DEIM as comparison methods; Figure 6 , Figure 7 and Figure 8 The qualitative test results and corresponding response heatmaps are shown on State-Air, AU-AIR, and VisDrone, respectively.
[0078] Figure 6Four representative challenging aerial remote sensing scenarios are presented, characterized by extremely small targets, large areas of background dominance, and complex structural textures. In wide-field-of-view scenarios containing a large number of tiny targets (row 1), YOLOv12 frequently misses detections, indicating its insufficient sensitivity to weak target cues that are overwhelmed by background structures. YOLOv13 improved recall in some scenarios, but still struggled to reliably locate small instances, especially when spatial details and semantic context were strongly coupled (rows 1 and 3). D-FINE could detect the main target, but tended to generate redundant bounding boxes and background-triggered responses when strong edges and textures were present (rows 2 and 4). DEIM showed more pronounced overactivation, with background structures such as roads and lane lines often being highlighted (rows 1, 2, and 4), leading to false detections and duplicate detections. In contrast, Fast-DETR produced more compact and stable detection results in all scenarios. Thanks to the spatial-context decoupled representation, the boundaries of small targets were clearer and the localization was more consistent, while redundant detections caused by structured background interference were significantly reduced. In addition, false detections occurring near the real target area were effectively suppressed, thereby improving the foreground-background separation effect. The response heatmap further showed that Fast-DETR concentrated activation in the real target area, avoiding unstable responses on the background-dominant structure, which is consistent with its decoupled representation learning and matchability-aware optimization design.
[0079] Figure 7 Qualitative results on AU-AIR are presented. This dataset includes small vehicles, dense traffic, varying viewpoints, and complex road layouts. Under these conditions, YOLOv12 and YOLOv13 are more prone to false negatives and less accurate localization, often producing redundant predictions in congested scenes (rows 1-3). D-FINE captures most salient instances, but duplicate detections and occasional false negatives still occur in dense small target areas, especially when high-contrast background textures dominate the scene (rows 1 and 2). DEIM tends to produce false positives driven by the background, with detection responses often extending to non-target areas such as road surfaces and surrounding structures. In contrast, Fast-DETR outputs cleaner: tighter bounding boxes and significantly fewer duplicate predictions. Even in dense traffic scenes, small vehicles can be located more accurately, and spurious responses around real targets are significantly reduced. This reflects more effective cross-scale feature aggregation and stronger suppression of blurred background responses during the optimization phase. Heatmaps show that Fast-DETR focuses activation on vehicle regions while suppressing responses to lane lines and road edges, demonstrating its stronger robustness to background interference and scale changes.
[0080] like Figure 8As shown, VisDrone simultaneously handles extremely dense small targets and highly cluttered background structures. YOLOv12 and YOLOv13 exhibit missed detections and loose localization issues under dense small target conditions, and generate a large number of redundant bounding boxes and false detections in areas dominated by strong textures or strong edges (lines 1, 3, and 4). D-FINE can detect major instances in some scenes, but still experiences missed detections and duplicate predictions in extremely dense areas, accompanied by background-triggered responses (lines 1-3). DEIM has a more significant tendency to false detections, and its predictions often cover high-response background areas, causing local confusion between the target and the surrounding background (lines 1, 2, and 4). In contrast, Fast-DETR can stably output cleaner predictions: redundant bounding boxes are significantly reduced, and the localization of dense small targets is more precise. In particular, false detections near real targets are effectively suppressed, making the separation between foreground targets and their immediate background clearer. The heatmap also further verifies that Fast-DETR emphasizes discriminative target regions and avoids overactivation on large areas of textured background. This is consistent with its matchability perception loss design: the design explicitly strengthens the supervision of difficult negative samples near the target and alleviates the noise gradient caused by ambiguous samples, thereby improving the detection robustness in dense drone scenarios.
[0081] Table 2 Comparison of computational complexity and inference speed 3) Operating efficiency: This invention further compares the computational complexity and inference efficiency of Fast-DETR with the above 12 representative detectors; to ensure fairness, all experiments were conducted on the same NVIDIA A100 (80GB VRAM) GPU.
[0082] As shown in Table 2, Fast-DETR exhibits the best efficiency profile among all comparison methods: it has the highest inference speed, reaching 34.60 FPS, while requiring only 16.13 GFLOPs, which is the lowest computational cost among all models; in addition, Fast-DETR contains only 7.88M parameters, making it the smallest model size among all comparison methods, including YOLO-based detectors and recent DETR variants.
[0083] Fast-DETR's high efficiency is primarily due to its self-attention-free encoder design, which avoids the secondary computational overhead of the Transformer encoder. By employing lightweight convolutional coding, Fast-DETR significantly reduces computational complexity while maintaining discriminative representations, thereby achieving faster, more energy-efficient inference and improved detection performance.
[0084] C. Ablation experiment.
[0085] 1) Structural Ablation: This invention conducts ablation experiments on the STATE-AIR dataset to evaluate the contributions of each key module (i.e., SCDM, DGFA, and MAL-Dynamic Loss (DMA-Loss)). The results are summarized in Table 3, where the baseline model (B) represents the network without any of the proposed components.
[0086] Table 3 Ablation experimental results of key components on the STATE-AIR dataset As shown in Table 3, simply introducing SCDM can increase AP from 49.4% to 52.3%, and also increase AP 75 The AP improved from 48.5% to 52.7%; consistent gains were achieved across small, medium, and large targets, indicating that it can learn more reliable multi-scale representations; using DGFA alone also brought a significant improvement in AP (49.4%→51.6%), and significantly reduced computational overhead (GFLOPs decreased from 24.84 to 13.45, and the number of parameters decreased from 10.18M to 7.50M), indicating that differential-guided cross-scale propagation can effectively enhance semantic aggregation with a lightweight budget; when using DMA-Loss alone, the AP further improved to 53.3%, with more significant improvements on small and medium targets, verifying that gradient redistribution based on matchability can stably optimize the process under localization uncertainty.
[0087] When any two components are combined, a significant complementary effect can be observed. SCDM+DGFA achieves an AP of 52.7% and improves AP. 75 While maintaining low computational complexity, this demonstrates that decoupled representations contribute to efficient cross-scale aggregation. SCDM+DMA-Loss achieved the highest AP (53.4%) and best APs (45.3%) across all "two-component" settings, highlighting the strong synergistic effect of representation decoupling and matchability alignment optimization in small target detection. DGFA+DMA-Loss also improved AP to 52.5% and achieved significant gains on large targets (AP... L =83.6%), indicating that efficient feature aggregation and localization-aware supervision can reinforce each other.
[0088] When all three are integrated, the complete model achieves more balanced performance across all scales, reaching AP=52.7% and AP=52.7%. 50 =93.2%, AP 75 =53.2%, while maintaining a moderate computational cost (16.13 GFLOPs and 7.88M parameters); in APs, AP M AP LThe continuous improvement validates the complementarity of SCDM, DGFA and DMA-Loss. Together they constitute an end-to-end "decoupling-alignment" framework, which synergistically improves the performance of UAV target detection in complex remote sensing scenarios from three levels: representation learning, feature aggregation and optimization strategy.
[0089] Table 4 Ablation experimental results on DMA-Loss on the STATE-AIR dataset 2) DMA-Loss Ablation: Table 4 reports the impact of DMA-Loss hyperparameters a and b on detection performance. Overall, the AP under different settings is basically stable, fluctuating only within a narrow range of 52.6%–53.3%, indicating that the loss is not sensitive to the choice of hyperparameters; increasing b usually brings a slight performance improvement, while changes in a mainly cause marginal fluctuations; the best results are achieved with a=1.0 and b=1.5, obtaining an AP of 53.3% after 200 epochs of training, so this configuration was used in all experiments.
[0090] D. Robustness testing.
[0091] To evaluate the robustness of Fast-DETR under real UAV imaging conditions, this invention conducts robustness tests on the STATE-AIR, AU-AIR, and VISDRONE datasets. A perturbation environment is constructed by simulating eight common image degradation types, including Gaussian noise, shot noise, impulse noise, snowfall, fog, brightness variations, contrast variations, and pixelation. Fast-DETR is compared with five representative low-altitude UAV target detectors: FBRT-YOLO, EDNet, UAV-DETR, D-FINE, and DEIM. Following a robustness evaluation protocol, robustness is quantified using three metrics: P represents performance on clean data; mPC represents average performance under various perturbations; and rPC represents relative performance consistency.
[0092] Table 5. Comparison Results of Robustness Tests As shown in Table 5, Fast-DETR achieves the highest AP(P) on clean data and also demonstrates stronger robustness on mPC and rPC, indicating that it possesses both strong detection capabilities and stable performance under diverse image degradation conditions; the same trend was observed on AU-AIR and VISDRONE; Figure 9 The qualitative results further demonstrate that Fast-DETR can maintain more stable detection output under different degradation scenarios, highlighting its robustness advantage in challenging UAV imaging environments.
[0093] E. Generalization experiment.
[0094] This invention further evaluates the generalization ability of Fast-DETR on the TT100K dataset; TT100K is a large-scale traffic sign detection benchmark with diverse scenarios, significant scale variations, and frequent background clutter; Table 6 presents the quantitative comparison results with 12 representative detectors, covering YOLO series, DETR-like methods, and two-stage detection methods.
[0095] Table 6 Quantitative comparisons on the TT100K dataset Fast-DETR achieved the best overall performance, with AP (64.3%) and AP (64.3%) respectively. 50 (85.0%), AP 75 (75.5%), APs (51.8%) and AP M (70.3%) reached the highest level; notably, compared to the strongest baseline, Fast-DETR improved by 0.4%–1.2% on most metrics, indicating its more robust detection capability for small- and medium-scale traffic signs under complex conditions; AP L The slight decrease (-6.1%) is mainly due to the limited number of large target samples in TT100K, which has little impact on the overall detection performance. The results show that Fast-DETR is not only applicable to UAV datasets, but also has good generalization and detection capabilities in urban driving environments.
[0096] In the representation stage, this invention introduces a Space-Context Feature Decoupling Module (SCDM), which explicitly separates the feature map into a context-aware branch and a spatially guided branch. The context-aware branch is used to enhance semantic modeling, while the spatially guided branch is used to preserve geometric details. This resolves the representational ambiguity between target identification and spatial localization within independent subspaces, reducing the mutual interference between semantic abstraction and structure preservation. In the aggregation stage, a Differential Guided Feature Aggregation Module (DGFA) is designed. By modeling inter-scale differences, cross-layer semantic propagation is controlled, achieving stable, directional global context fusion without self-attention. In the training stage, a Dynamic Matchability-Aware Loss (DMA-Loss) is proposed. Gradients are adaptively allocated between the target and its surrounding background based on the localization quality, strengthening the discrimination boundary and stabilizing training under dense, small target distributions. Extensive experiments on multiple low-altitude UAV benchmark datasets demonstrate that Fast-DETR achieves state-of-the-art detection performance, especially with significant advantages on the smallest targets, while achieving the fastest inference speed and lowest computational cost among comparable UAV detectors.
[0097] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A target detection method for visible light images from unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Acquire visible light images captured by a drone; A visible light image is input into a pre-trained target detection model, which includes a backbone network, a feature decoupling module, and a feature aggregation module. Multi-scale feature maps of visible light images are extracted through a backbone network; differential Gaussian operations are performed on the feature maps of each scale through the spatial guidance branch of the feature decoupling module to obtain high-frequency residual features representing edge texture; spatial attention weight maps of high-frequency residual features are generated using a spatial attention mechanism; and spatial attention weight maps are multiplied element-wise with the feature maps of the corresponding scale to obtain spatial edge texture features. The context-aware branch of the feature decoupling module uses large kernel convolution to aggregate context information in the local region of the feature map at each scale to obtain region aggregated features. Global average pooling is used to obtain the channel semantic statistical vector of the feature map at the corresponding scale. The region aggregated features and the channel semantic statistical vector are added and fused element by element to obtain the semantic attention weight map. The semantic attention weight map is multiplied element by element with the feature map at the corresponding scale to obtain the semantic context features. Then, the spatial edge texture features corresponding to each scale feature map are added and fused with the semantic context features element by element to obtain the decoupling enhancement features corresponding to each scale; The feature aggregation module processes adjacent scale pairs step by step along the feature pyramid, extracts the feature difference between the decoupled enhancement features of each adjacent scale pair as a guiding signal, generates channel modulation weights through a lightweight adaptive controller based on the guiding signal, and uses the channel modulation weights to perform weighted fusion of the decoupled enhancement features of the scale pair to obtain the intermediate aggregated features of the scale pair. The intermediate aggregated features are used as the new current scale and continue to perform differential guided aggregation with the decoupled enhancement features of the next adjacent scale to obtain the final aggregated features. Targets in visible light images are detected based on the final aggregated features.
2. The target detection method for visible light images of UAVs according to claim 1, characterized in that, The acquisition of the spatial edge texture features includes: The multi-scale feature map of the visible light image is as follows: ,make and For channel-by-channel Gaussian kernels applied independently, and whose standard deviations satisfy... The high-frequency response of capturing edges and textures is represented as: Where: represents the convolution operation; Two channel-wise Gaussian kernels and Using a single Gaussian kernel By approximation, the high-frequency residual calculation is expressed as: After obtaining the high-frequency residual, use Convolution for local structural refinement and noise suppression is represented as follows: After local structural refinement and noise suppression, pointwise linear dimensionality reduction-recovery of the PWLR bottleneck is used to adaptively weight spatial location and channels to obtain a spatial attention weight map. Let the dimensionality reduction ratio be... , ,but: in: , Indicates learnable Convolution kernel; Represents the Sigmoid function; Applying the spatial attention weight map to the input features yields spatial edge texture features, represented as: in: This indicates element-wise multiplication.
3. The target detection method for visible light images of UAVs according to claim 2, characterized in that, The acquisition of the semantic context features includes: Using large-kernel depthwise convolutions, contextual information is aggregated within local regions expanded at each scale feature map to obtain region aggregated features, represented as: in: For learnable Convolution kernel; Global channel-level semantic priors are obtained through global average pooling, resulting in channel semantic statistical vectors, represented as follows: in: This represents the average activation of each channel in the spatial domain; The region aggregation features and channel semantic statistical vectors are transformed using the PWLR bottleneck and expressed as follows: The transformed features are fused element-wise to obtain the semantic attention weight map, represented as: Applying the semantic attention weight map to the original features yields semantic context features, represented as follows: in: This indicates element-wise multiplication.
4. The target detection method for visible light images of UAVs according to claim 3, characterized in that, The acquisition of the intermediate aggregation features includes: make , The decoupling enhancement features at two scales are represented by projecting both onto a shared low-dimensional subspace using low-rank convolution, as follows: in: ; Differential Feature Map The differential feature map is used as a guiding signal for feature aggregation: In top-down propagation, i.e., higher levels to lower level During the dissemination, This indicates the missing fine-grained details in high-level features, guiding the module to selectively inject low-level spatial information; In bottom-up aggregation, i.e., the lower layer To the upper levels During aggregation, Capture supplementary semantic abstractions in low-level features to enable selective reinforcement of high-level semantic cues; The channel-level global descriptor is obtained through explicit spatial averaging, and is represented as follows: in: ; The global descriptor is obtained by concatenating the three elements. The global descriptor is then input into the lightweight two-layer controller to generate channel modulation weights, represented as: in: Corresponding to and Channel weights; Modulating and fusing the features is represented as: Local refinement is achieved using convolution and residual connections are introduced to obtain intermediate aggregated features, represented as follows: in: These are learnable residual scaling parameters.
5. A target detection method for visible light images of unmanned aerial vehicles according to claim 1, characterized in that, During training, the target detection model employs a Dynamic Matchability-Aware Loss (DMA-Loss), which includes: The first The loss for each query is defined as: in: The predicted prospect probability; q i The focus index represents the positioning quality measured by the intersection-union ratio (IU). Based on the dynamic determination of positioning quality, it is expressed as: in: Control the sensitivity to positioning errors; Set the minimum focus intensity; Loss through location quality q i The relevant gradient modulation for two-dimensional classification is represented as follows: when hour, Smaller The losses were mainly caused by Dominant, used to enhance foreground supervision of reliable positive samples; when hour, Increase inhibition And enhance the background-related terms to weaken the supervision strength of unreliable positive samples, in order to reduce the noise gradient caused by low-quality matching; when hour, Decrease Amplification A stronger penalty is applied to high-confidence false positives that are close to the real target, which is used to distinguish the foreground from the background; when hour, Increase inhibition This is used to focus training on more informative samples.
6. The target detection method for visible light images of unmanned aerial vehicles according to claim 1, characterized in that, The target detection model adopts an encoder-decoder architecture, with the HGNetv2 backbone network as the backbone network.
7. A target detection device for visible light images from unmanned aerial vehicles (UAVs), characterized in that, include: Image module, used to acquire visible light images captured by the drone; The detection module is used to input visible light images into a pre-trained target detection model, which includes a backbone network, a feature decoupling module, and a feature aggregation module. Multi-scale feature maps of visible light images are extracted through a backbone network; differential Gaussian operations are performed on the feature maps of each scale through the spatial guidance branch of the feature decoupling module to obtain high-frequency residual features representing edge texture; spatial attention weight maps of high-frequency residual features are generated using a spatial attention mechanism; and spatial attention weight maps are multiplied element-wise with the feature maps of the corresponding scale to obtain spatial edge texture features. The context-aware branch of the feature decoupling module uses large-kernel depth convolution to aggregate context information in the local region of the feature map at each scale to obtain region aggregated features. Global average pooling is used to obtain the channel semantic statistical vector of the feature map at the corresponding scale. The region aggregated features and the channel semantic statistical vector are added and fused element by element to obtain the semantic attention weight map. The semantic attention weight map is multiplied element by element with the feature map at the corresponding scale to obtain the semantic context features. Then, the spatial edge texture features corresponding to each scale feature map are added and fused with the semantic context features element by element to obtain the decoupling enhancement features corresponding to each scale; The feature aggregation module processes adjacent scale pairs step by step along the feature pyramid, extracts the feature difference between the decoupled enhancement features of each adjacent scale pair as a guiding signal, generates channel modulation weights through a lightweight adaptive controller based on the guiding signal, and uses the channel modulation weights to perform weighted fusion of the decoupled enhancement features of the scale pair to obtain the intermediate aggregated features of the scale pair. The intermediate aggregated features are used as the new current scale and continue to perform differential guided aggregation with the decoupled enhancement features of the next adjacent scale to obtain the final aggregated features. Targets in visible light images are detected based on the final aggregated features.
8. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the target detection method for visible light images of unmanned aerial vehicles as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the steps of a target detection method for visible light images of unmanned aerial vehicles as described in any one of claims 1 to 6.