Continuous poultry monitoring requires accurate localization under posture variation, partial occlusion, repeated housing structures, and edge-computing constraints, all of which can weaken cross-scale representation and increase bounding-box uncertainty. We adapt the real-time detection transformer (RT-DETR) with scale-aware dynamic weighting for complete intersection over union (SADW-CIoU), a training-time regression mechanism, and a complementary Coordinate Attention neck (CA_Neck) for feature refinement. SADW reweights the matched final-decoder CIoU term after assignment using detached scale and quality factors; CIoU retains its regression gradient, while the Hungarian matching cost and auxiliary, encoder-auxiliary, and denoising branches remain unchanged. On a 2090-image poultry house dataset, three-seed controlled experiments increased the mean average precision over IoU thresholds 0.50−0.95 (mAP _50:95 ) from 0.7853± 0.0025 for the standardized RT-DETR-L baseline to 0.8230± 0.0020 with SADW and 0.8297± 0.0015 with CA_Neck+SADW. Controlled regression and attention comparisons, sensitivity tests, and assignment diagnostics delimit the contributions of the CIoU/scaling foundation and adaptive localization weighting, with an approximately additive contribution from CA_Neck; no universal superiority or formal significance is inferred. SADW adds no inference structure: RTX 4090 forward latency was 6.80 ms for the baseline and 6.82 ms for SADW-only, while the combined model retained 143.06 FPS. Archived TensorRT measurements on a Jetson AGX Orin provide separate embedded engine-level evidence. Overall, the method improves strict-IoU localization while preserving real-time inference within the evaluated poultry monitoring conditions.
Object detection in unmanned aerial vehicle (UAV) imagery is challenging because targets are often small and densely distributed. Complex backgrounds and real-time requirements make the task even harder. Existing detectors often fail to preserve discriminative features for small objects. They also have difficulty modeling useful context efficiently in crowded scenes. As a result, it is hard to achieve both high accuracy and low computational cost. To address these issues, we propose MS-DETR, a real-time UAV object detector based on Real-Time Detection Transformer (RT-DETR). The method follows a refine-then-fuse design. First, a multi-domain feature learning backbone enhances small-object representations by combining spatial, channel, and frequency cues. Then, an efficient sparse attentive fusion scheme improves contextual modeling and multi-scale feature aggregation with low overhead. In this way, MS-DETR improves fine-grained perception for small objects and reduces redundant background responses, while keeping high efficiency. Extensive experiments on the VisDrone, HazyDet, and CODrone datasets show that MS-DETR achieves a competitive accuracy-efficiency trade-off. Compared with the RT-DETR baseline, it improves mean average precision over intersection-over-union thresholds from 0.50 to 0.95 (mAP50:95) by 2.4–4.3 points on different benchmarks. At the same time, it reduces the number of parameters by about 33
Small objects in UAV imagery usually produce weak detection responses. Existing methods lack a small-object weight calibration mechanism across multiple detection stages, making it difficult to preserve effective small-object responses and fully optimize subtle localization errors. To address these issues, we propose MAWC-DETR, a UAV small-object detection framework based on multi-stage adaptive weighted calibration. In the feature extraction stage, cross-scale weighted convolution (CSWC) is proposed to address the equal-weight modeling limitation across spatial positions. It adaptively enhances local responses of small objects while enabling lightweight and discriminative feature extraction. In the feature fusion stage, a recalibration attention network (RCAN) recalibrates small-object feature weights through detail recovery and semantic modeling, compensating for feature loss caused by sampling and fusion. In the regression supervision stage, WMFIoU is constructed to break the equal-penalty pattern for samples and errors. It adaptively focuses on difficult small objects and subtle localization deviations, improving bounding-box regression accuracy. By jointly optimizing feature extraction, fusion, and regression supervision, MAWC-DETR enhances effective small-object representation under lightweight constraints. Experiments on VisDrone2019-DET show that MAWC-DETR achieves 52.0
In power line inspections, accurate semantic segmentation is a core link in ensuring the safe and stable operation of power transmission systems, with significant engineering value for UAV-based maintenance and reducing fault-induced economic losses. However, this task is invariably confronted with complex challenges, which encompass complex backgrounds, delicate slender structures of power lines, the contradiction between high accuracy and real-time performance, and pronounced class imbalance between power lines and background. To address these issues, this paper proposes the WCM-DeepLab model for real-time power line segmentation in complex environments. Specifically, a Wavelet-based Convolutional Modulation Network (WCMNet) is built as the backbone, including a BasicBlock fusing local and global features via convolutional modulation and a Wavelet Attention Downsample (WAD) module compensating for detail loss. Further, a lightweight Strip Atrous Spatial Pyramid Pooling (SASPP) module enhances accuracy through multi-scale information and asymmetric convolutional attention. Additionally, a hybrid loss function combining cross-entropy and Dice loss mitigates class imbalance. Experimental results on the TTPLA dataset show the model outperforms advanced methods, achieving a mean Intersection over Union (mIoU) of 78.31
Real-time object detection for unmanned aerial vehicles (UAVs) in foggy environments remains challenging due to visibility degradation, scale-dependent feature attenuation, and the limited computational resources onboard. Existing methods often rely on computationally intensive dehazing–detection pipelines or adopt fixed receptive fields, which are insufficient for dynamic fog scenarios and small or distant targets. To address these limitations, we propose DSC-YOLOv8s, a lightweight detection framework for UAVs in foggy environments. It integrates Dynamic Multi-Scale Gated Convolution (DMS-GC) to adaptively enhance features and a Dual-Dimensional Scale-aware Cross-Attention (D2-SCA) module to efficiently model spatial and channel dependencies. DMS-GC adjusts receptive fields across multiple scales to restore fog-degraded features with minimal computation, while D2-SCA strengthens cross-dimensional dependencies to improve detection of small and motion-blurred targets. An auxiliary repulsion-based loss refines localization in dense scenes. Experiments on HazyDet and RTTS datasets show that DSC-YOLOv8s improves mAP50 by 8.6
