Effect of initial learning rate (LR) on mAP.

The rapid advancement of Unmanned Aerial Vehicle (UAV) technology has facilitated dynamic, high-resolution remote sensing, significantly benefiting applications in agriculture, forestry, urban planning, and disaster management. However, detecting small objects in UAV imagery remains challenging due to severe scale variations and environmental complexities. While traditional detection methods and even many advanced YOLO variants achieve reasonable performance, they often either incur high computational costs or fail to preserve the fine-grained features essential for reliably detecting extremely small targets. To overcome these limitations, we propose SRD-YOLOv5, an enhanced version of the lightweight YOLOv5n model, distinguished by its novel multi-scale feature fusion framework. Our approach introduces two innovative modules: the Scale Sequence Feature Fusion Module (SSFF) and the Multi-Scale Feature Extraction Module (MSFE), which collaboratively capture global contextual information and preserve detailed semantic cues that are typically lost in conventional fusion techniques. Furthermore, we incorporate an Extremely Small Target Detection Layer (ESTDL) specifically designed to retain high-resolution features for micro-scale object detection. Additionally, the implementation of a Decoupled Head, which independently processes regression and classification tasks, further optimizes the detection of small targets by reducing task conflicts and improving localization precision. Experimental results demonstrate that SRD-YOLOv5 outperforms existing methods in detecting small targets within UAV remote sensing images. It achieves higher accuracy while maintaining low computational demands, making it suitable for real-time applications in UAV remote sensing.