Rethinking Dual-Stream Super-Resolution for Enhancing Remote Sensing Object Detection

Abstract

Detecting small instances within complex backgrounds presents significant challenges for Remote Sensing Object Detection (RSOD). While complex deep networks can enhance feature representation, they often lead to considerable computational burdens. Previous research has proposed dual-stream learning to improve the detection capabilities of compact models. We have re-evaluated existing dual-stream learning frameworks and identified their limitations in focusing on small objects. To address this issue, we propose a Dual-Stream Object Detection (DSOD) framework, which incorporates a Feature Fusion Guidance Module (FFGM). Specifically, by integrating features from the two task streams under dual-task supervision, DSOD directs the super-resolution process towards instance-related regions. This approach enhances feature extraction with object instance-specific details, significantly improving RSOD accuracy without introducing additional computational overhead. The effectiveness of DSOD has been validated across three models on three publicly available datasets. Furthermore, ablation studies highlight the importance of DSOD in optimizing RSOD performance while maintaining computational efficiency.