CAF-YOLO: a UAV small object detection network based on feature complementary perception and adaptive alignment fusion

Abstract

Object detection in UAV aerial imagery is a key image-based visual measurement task for extracting object categories, image-plane locations, spatial extents, and distribution information from airborne sensor data. However, it remains challenging because of severe scale variation, complex background noise, motion blur, and extremely small pixel footprints of distant objects. To address these limitations, this study proposes CAF-YOLO, a Complementary Alignment and Fusion Object Detection Model based on YOLO11. The model improves UAV-based visual measurement reliability through three designs. First, Progressive Multi-scale Attention Convolution (PMAConv) reduces spatial information loss for small objects during downsampling. Second, Adaptive Residual Alignment Fusion Module (ARAFM) is embedded in the feature fusion network, and Triple-Efficient Attention Module (TEAM) is placed before the detection head. They enable adaptive cross-scale feature fusion and background-interference suppression, enhancing small-object detection and reducing false positives and false negatives in complex aerial scenes. Third, Log-Rational IoU (LRIoU), incorporating rational algebraic projection, a Fisher–Rao-inspired logarithmic scale measure, and exponential decoupling, improves bounding-box localization stability and sensitivity to small-object scale deviations. Experiments on VisDrone2019 demonstrate competitive overall detection and localization performance, with the main gains observed in mAP50 and mAP50-95 while retaining a moderate parameter count. With 10.13 M parameters, CAF-YOLO achieves 54.6% Precision, 38.0% Recall, 46.0% mAP50, and 28.1% mAP50-95. Supplementary transfer-adaptation experiments on CARPK further indicate its effectiveness in another aerial-view vehicle-detection scenario, supporting UAV-based visual measurement applications.