KBT-YOLO: A Lightweight and Efficient Small Object Detector for UAV Aerial Imagery

Abstract

Object detection in Unmanned Aerial Vehicle (UAV) aerial imagery frequently suffers from missed detections and false alarms due to the extremely small scale of objects, dense distributions, and highly complex backgrounds. To mitigate these challenges, this paper proposes a novel lightweight small object detection method tailored for UAV aerial images. First, a dynamic convolution module is introduced to replace the standard convolutions within the C3k2 module, which simultaneously enhances feature extraction capabilities and effectively reduces computational redundancy. Second, a Bidirectional Feature Pyramid Network (BiFPN) architecture is incorporated to dynamically balance feature interactions via learnable weights, thereby improving local feature capturing and the effectiveness of multi-scale fusion. Finally, a task-aligned dynamic detection head based on a dual-branch structure is designed. By integrating task decomposition and alignment mechanisms with dynamic convolutions, this detection head significantly boosts both detection accuracy and stability. Extensive experiments on the DIOR UAV aerial dataset demonstrate that the proposed method achieves leading detection accuracy while substantially minimizing the parameter count to merely 1.85M. This research offers a promising new technical paradigm for lightweight object detection in UAV applications.