SEFDetect: A Spatially Enhanced Fusion Detector for Aerial Imagery

Unmanned Aerial Vehicle (UAV) technology has significantly advanced, finding extensive applications in traffic planning, agricultural production, and ecological protection, thereby promoting development in these fields. However, object detection in UAV aerial imagery continues to face significant challenges, including dense small objects, disruption from complex backgrounds, and substantial variations in object sizes. To reduce the impact of these challenges on recognition effectiveness, this paper proposes an object detection model, SEFDetect. Firstly, the Spatial Kernel Selection Module (SKSM) is designed to dynamically adjust the receptive field, so that the model can focus on key features in the image. Secondly, the Channel Conversion Attention Module (CCAM) is introduced, which enhances the robustness of the algorithm by extracting detailed visual information and improving the spatial perception of the network. Thirdly, an adaptive feature fusion structure is constructed to explore the contributions of different resolution feature maps. Finally, to ensure a fair evaluation of SEFDetect, extensive experiments were conducted on the VisDrone2021 and PASCAL VOC datasets under consistent conditions.