Star-PMFI: Star-attention and pyramid multi-scale feature integration network for small object detection in drone imagery

With their high flexibility and cost-effectiveness, Unmanned Aerial Vehicle (UAV) plays a crucial role in target detection and are widely used in military, rescue, and traffic surveillance scenarios. However, due to its particular aerial viewpoint, UAV images contain many small and densely distributed targets, which poses a severe challenge for accurate detection. In this study, we propose a novel UAV target detection model, Star-PMFI, consisting of the Star-Attention (Star-A) backbone network and the Pyramid Multi-scale Feature Integration (PMFI) neck. The Star-A utilizes the star operation and attention mechanism to extract the rich features, and the PMFI module performs the initial integration of features through the pyramid structure, followed by in-depth feature interaction. First, the model extracts multi-scale features using Star-A, which skillfully combines the star operation and attention mechanism to capture an extensive range of contextual information. Second, PMFI initially integrates the features through the pyramid structure, followed by deep feature interaction to realize cross-scale and cross-level information fusion. Finally, the model employs six detection heads, each responsible for target detection at different scales or features, to enhance small target detection capability. The experimental results show that the Star-PMFI model performs excellently on multiple datasets. On VisDrone and UAVDT datasets, $mAP@0.5:0.95$ reaches 28.7% and 84.0%, respectively. Our code is available at: https://github.com/yangwygithub/PaperCode/tree/main/WenyuanYang_Star-PMFI_UAV