Integrated Traditional Chinese and Western Medicine for Chronic Spontaneous Urticaria: Protocol for a Trials Within Cohorts Study.

With the rapid development of unmanned aerial vehicle (UAV) technology, UAV systems equipped with visual sensors and artificial intelligence algorithms have been widely used in urban management, traffic monitoring, agricultural monitoring, emergency rescue [], and other fields. UAV aerial images are characterized by large variations in viewpoint, complex and diverse backgrounds, and significant differences in target scale. In particular, small targets in images (e.g., pedestrians and vehicles) are difficult to detect due to their low resolution and weak feature representations [], which has become a major challenge in object detection research. Although traditional object detection methods, such as YOLO, R-CNN [], and Faster R-CNN [], have achieved good performance in some scenarios, they still suffer from insufficient accuracy and high miss rates for small objects [] in UAV aerial images, especially when handling multi-scale objects and complex backgrounds. This remains a common challenge in small-object detection research. Currently, many researchers in the field of UAV object detection are improving YOLO-series algorithms to address the difficulties of detecting objects in aerial images and to enhance detection accuracy. For example, Zhu et al. [] proposed the TPH-YOLOv5 algorithm, which adds a small-object detection head and a Transformer prediction head (TPH), and incorporates a convolutional block attention module (CBAM) to improve object localization accuracy in high-density scenes. Similarly, Lin et al. [] proposed the HTH-YOLOv5 algorithm, which significantly improves the detection accuracy of small objects in UAV images by introducing a Hybrid Transformer detection head (HTH) alongside CBAM. In recent years, the DETR algorithm [], based on the Transformer architecture, has gained popularity in object detection due to the advantages of its attention mechanism in capturing semantic information. However, the DETR model still faces challenges, including high complexity, difficulty in achieving training convergence, and unsatisfactory performance in small-object detection []. To address these issues, Zhu et al. [] proposed Deformable DETR, which utilizes a deformable attention mechanism to extract multi-scale features, thereby reducing model complexity and accelerating convergence. Zhao et al. [] introduced RT-DETR, which significantly enhances detection accuracy and real-time performance through an efficient hybrid encoder and IoU-aware query selection, laying a theoretical foundation for applying DETR models to real-time object detection in UAV applications. Subsequently, researchers have made several improvements to RT-DETR: Zhou et al. [] proposed the UAV-DETR algorithm, which effectively enhances the detection accuracy of small objects in UAV images by introducing a Channel-Aware Module (CAS), a Scale-Optimized Enhanced Pyramid (SOEP) module, and a Contextual Spatial Alignment Module (CSAM). However, this method still faces limitations when processing low-resolution images. Su et al. [] introduced the DRT-DETR algorithm, which significantly improves detection accuracy and computational efficiency by incorporating a Fast Multi-scale Attention Feature Extraction Module (Faster-EMA) and a Weighted Bidirectional Cross-scale Feature Fusion Module (Bi-CCFM). Nevertheless, this approach still carries the risk of missed detections when handling small targets in complex backgrounds. Shen et al. [] proposed the TinyDef-DETR algorithm, which enhances the model’s ability to detect small objects in power line defect inspection by introducing a lossless downsampling module, edge-enhanced convolution, a cross-stage dual-domain multi-scale attention module, and a focus-aware regression loss function. However, the model is relatively sensitive to background noise and may exhibit false detections or missed detections in natural environments or cluttered scenes. Tong et al. [] proposed the ACD-DETR algorithm, which improves small-object detection performance by incorporating a Multi-scale Edge Enhancement Fusion Module (MSEFM), a Global Boundary Calibration module (OG-BC), and a Dynamic Position Bias Attention module (DPB-AIFI). Despite these improvements, the model’s stability and accuracy still require further enhancement when handling complex background scenarios.