A small object detection method with context information for high altitude images

Abstract

Detection of small objects stands as a pivotal and difficult task because of their low resolution and lack of visualization features. Though achieving some promising results, recent detection methods utilize the context information insufficiently, leading to inadequate small object feature representation and increasing the misdetection and omission rates. We propose a method named Context Information Enhancement YOLO(CIE-YOLO) for small object detection. CIE-YOLO mainly includes a Context Reinforcement Module(CRM), a Channel Spatial Joint Attention(CSJA) module, and a Pixel Feature Enhancement Module(PFEM). The CRM module extracts and enhances the context information to mitigate the confusion between small objects and the background in the network. Then CSJA suppresses the background noise to highlight important small object features. Finally, PFEM reduces the small object feature losses in up-sampling via feature enhancement and pixel resolution enhancement. The effectiveness of the proposed CIE-YOLO in small object detection is demonstrated by extensive experiments.