TOReN: Tiny Object Recognition Network in Remote Sensing Imagery

The surge in high-resolution satellite and aerial imagery has created urgent demands for effective tiny object recognition, yet conventional deep learning frameworks struggle with objects smaller than Formula: see text pixels due to spatial information loss during network downsampling. This paper presents TOReN, a Tiny Object Recognition Network framework addressing the fundamental challenge where advanced sensor capabilities enable unprecedented Earth observation while rendering critical tiny targets nearly imperceptible to existing recognition models. The proposed TOReN integrates three key components: 1) an enhanced backbone network featuring higher-resolution feature preservation with adaptive deformable convolutions to maintain spatial integrity; 2) a DeNoising Feature Pyramid Network (DN-FPN) employing contrastive learning to mitigate semantic and geometric distortions in multiscale feature fusion; and 3) a Tiny Object Recognition and Fusion Head (TORF-Head) incorporating a specialized recognition branch with channel attention and fusion mechanisms based on object scale. Experiments conducted on the AI-TOD-v2 dataset demonstrate significant performance improvements for tiny object recognition, with TOReN achieving 28.5% average precision compared to the baseline FCOS detector’s 13.2% average precision (a 116% relative improvement).