Scale-normalized YOLO for small-object detection in large-oblique-angle UAV imagery

Abstract

Fixed-wing unmanned aerial vehicles (UAVs) have been widely employed in remote sensing inspection, disaster assessment, traffic surveillance, and ground target recognition because of their long endurance, wide-area coverage, and high imaging efficiency. However, high-resolution imagery captured under large oblique viewing angles is subject to severe perspective distortion and spatially varying ground sampling distance (GSD), resulting in substantial intra-class variations in object scale, shape, and texture. These factors significantly degrade detection performance, particularly for small vehicle-like targets in cluttered backgrounds. Moreover, directly resizing large-format images to the fixed input resolution required by common detection networks inevitably compresses small objects and causes the loss of discriminative details. To address these issues, this paper proposes a scale-normalized object detection method for largeoblique- angle fixed-wing UAV imagery based on an enhanced YOLO11 framework. First, geometric rectification is introduced to alleviate perspective-induced distortions and reduce the spatial inconsistency of object appearance. Second, a fixed-GSD-based local tiling and scale normalization strategy is developed to transform large high-resolution images into uniformly scaled sub-images suitable for network input. The influence of different GSD settings on detection accuracy and inference efficiency is further investigated to determine an appropriate operating scale. Third, an improved YOLO11s detector is constructed by incorporating lightweight re-parameterized convolution, attention-guided feature enhancement, and detection head adaptation, thereby strengthening the representation capability for small targets under complex background conditions. Experimental results demonstrate that the proposed method effectively mitigates the scale inconsistency inherent in large-oblique-angle imagery and improves detection accuracy while maintaining favorable computational efficiency. Compared with the baseline model, the proposed approach achieves consistent gains in mAP@0.5, recall, and small-object detection performance, confirming its effectiveness for fixed-wing UAV large-obliqueangle target detection tasks.