Small targets in remote sensing images suffer from degraded texture and edges due to network pooling, with further challenges from scale variations and arbitrary orientations. SuperYOLO has two key limitations: its pixel-level symmetric Multi-modal Fusion (MF) module fails to distinguish the channel-wise semantic contributions to small targets; and its CSP modules lack explicit spatial coordinate encoding, leading to feature redundancy and localization deviations. Two improvements in this paper: ① replacing MF with a Multi-scale Self-Attention Aggregation (MSAA) module to enhance key features (e.g., vehicle contours, thermal signals) via channel weight optimization; ② integrating Coordinate Attention (CA) into CSP modules to strengthen position encoding through spatial-feature interaction, improving spatial discrimination under complex backgrounds. 78.9% mAP50 on VEDAI (3.8% higher than SuperYOLO, with trucks/tractors up +18.9%/+5.8%); 61.94 mAP50 on AI-TOD (surpassing FFCA-YOLO/SuperYOLO); 72.83 mAP50 on DOTA (exceeding counterparts). Ablation studies confirm MSAA+CA boosts generalization.
Qi et al. (Sun,) studied this question.