A Unified Object Detection Method in Drone View with Degradation-Aware and Domain Adaptive Modeling

Reliable object detection in drone views is of great significance for security, surveillance, and environmental monitoring, yet it remains severely challenged by adverse weather and domain shifts. Existing methods struggle with three major obstacles: (1) the significant distribution shift between clean and degraded samples under diverse weather conditions prevents models from robustly capturing intrinsic object representations; (2) drone views often exhibit small-scale and low-resolution targets, where even minor image degradations can drastically impair detection performance; and (3) most prior methods lack a unified and effective all-weather detection framework. To this end, a unified object detection method with degradation-aware and domain adaptive modeling is proposed. First, we design a degradation-aware module (DAM) that leverages amplitude characteristics in the frequency domain to explicitly model degradation patterns, enabling the detector to perceive and adapt to various types of image quality deterioration. Second, we propose a domain-aware attention based restoration expert system (DA-RES). It disentangles shared and domain-specific representations through a combination of domain-shared and domain-specific encoders, which effectively suppresses category-irrelevant information while enhancing domain-specific useful cues. Finally, through embedding the degradation patterns identified by DAM into the target domain encoder, DA-RES performs multi-scale feature restoration guided by degradation priors, thereby strengthening the downstream detection module against complex adverse environments. Extensive experiments on multiple drone-view benchmarks demonstrate that the proposed framework achieves robust and unified detection performance under all-weather scenarios. In particular, our approach delivers substantial improvements over state-of-the-art methods in challenging degraded conditions.