YOLP: You only Look Patches for Object Detection in UAV Remote Sensing Imagery

While object detection from unmanned aerial vehicles (UAVs) is vital for numerous applications, it confronts two major challenges: detecting small targets with limited effective pixels requires substantial computational resources, and significant feature extraction capacity is wasted on nontarget background regions because the objects are usually sparsely distributed and locally clustered. To improve the detector's robustness and computational efficiency, we propose CFFM-YOLO, an enhanced YOLOv8 detector incorporating a lightweight Cross-scale Feature Fusion Module (CFFM) for dynamic feature fusion across hierarchical feature levels. This module uses dynamic convolutions, scale attention, and channel attention to refine features across spatial, scale, and channel dimensions, boosting efficiency without compromising detection accuracy. For detecting small, non-uniformly distributed objects in complex image backgrounds, methods that employ extra networks for cropping image local regions suffer from high overhead, while increasing input resolution is wasteful due to extensive processing of irrelevant background regions. Building upon CFFM-YOLO with high-resolution inputs, we propose YOLP, which introduces a Feature Patch Slicing Module (FPSM) for selective feature-level cropping rather than processing the entire high-resolution image. The FPSM leverages backbone features to predict an object heatmap, generates size-fixed feature patches, and performs subsequent object detection. FPSM avoids redundant feature extraction caused by extra networks and skips redundant computation on background regions. On the VisDrone and UAVDT benchmarks, the proposed YOLP demonstrates superior detection performance, achieving state-of-the-art m A P scores of 3 8. 2% and 2 5. 5%, respectively. This result substantiates its effectiveness and generalization capability in UAV remote sensing object detection.