Precision Aircraft Detection via Synergistic Super-Resolution and Oriented Rotation Convolution

Precision detection of aircraft in aerial imagery faces two fundamental challenges: (1) minute target sizes often result in blurred or incomplete boundaries that hinder feature extraction, and (2) arbitrary in-plane rotations cause severe performance degradation in conventional object detectors designed for axis-aligned targets. These limitations lead to unreliable detection of small aircraft and frequent misalignment of bounding boxes for rotated instances. To address these issues, we propose an integrated framework that combines domainspecific super-resolution with rotation-adaptive feature learning. Our approach employs a multi-scale holistic attention network (MS-HAN) specifically optimized to reconstruct aircraft boundary details from low-resolution inputs while suppressing background interference through altitude-guided attention. Furthermore, we enhance the Dynamic Oriented Rotation Convolution (DORC) module with dynamic kernel compensation, which maintains feature integrity during rotation transformation while preserving computational efficiency. The unified framework demonstrates superior capability in detecting both small and arbitrarily rotated aircraft compared to existing solutions.