Enhanced Tiny Object Detection in Aerial Images

Our research addresses challenges posed by numerous tiny objects in aerial images, impacting object detection methods. These challenges stem from scale variation, smaller size, and lower resolution of tiny objects compared to medium and large ones, making feature extraction difficult. Common downsampling methods and the YOLOv5 detection head combining classification and regression tasks can lead to valuable feature loss. To address these issues, we propose an enhanced YOLOv5 framework. Our model introduces three key components: the Depth-Wise-Reshaping (DWR) module for downsampling, the Multi-branch Deep Stripe Attention (MDSA) module for multi-scale contextual enhancement, and the Decoupled-Head (DH) module for independent detection and regression tasks. By integrating these components, our model effectively reduces feature loss in tiny object detection, achieving outstanding results on three aerial datasets: AI-TOD, VisDrone2019, and RSOD. Comparative evaluations against baselines and other detection models demonstrate the superior performance of our approach.