FDA-DETR: A frequency-aware DETR with dynamic query and adaptive multi-task optimization for oriented small object detection

Abstract

Oriented small object detection remains a challenging problem in computer vision, largely due to the weak feature representation and high computational cost of existing detection Transformer (DETR)-based detectors. To address these issues, this work presents Frequency Domain Awareness Detection Transformer (FDA-DETR), an end-to-end framework that enhances both accuracy and efficiency for oriented small object detection. The core of FDA-DETR lies in its multi-scale frequency domain enhancement, which amplifies high-frequency details crucial for small object discrimination. And by introducing a density-aware dynamic query mechanism, the model further adapts computational resource allocation to object density and orientation, improving detection in complex scenes. To balance global context and local detail, a multi-granularity attention fusion module is incorporated, while an adaptive multi-task loss based on Bayesian uncertainty enables dynamic optimization across multiple objectives. Experiments on public datasets show that FDA-DETR achieves higher detection accuracy and faster inference speed compared to existing DETR-based methods, particularly for small and densely distributed objects. These results, supported by theoretical analysis and ablation studies, highlight the effectiveness and synergy of the proposed modules. FDA-DETR thus provides a robust solution for oriented small object detection and offers new perspectives for future research on feature learning and attention mechanisms.