Small object detection algorithm based on improved YOLOv12n
TL;DR
An improved algorithm based on YOLOv12n is proposed, with a progressive and complementary design that sequentially performs detail enhancement, cross-layer fusion optimization, and final detection, confirming its stability and efficiency for small‐object detection.
Abstract
Small object detection is challenging due to insufficient feature information and strong background interference. This paper proposes an improved algorithm based on YOLOv12n, with a progressive and complementary design that sequentially performs detail enhancement, cross-layer fusion optimization, and final detection. First, a high‐pass multi‐dimensional collaborative attention (HPMCA) module embedded in the backbone uses discrete cosine transform high‐pass filtering and multi‐dimensional attention to extract edge and texture details, strengthening subtle features in shallow and middle layers to compensate for inherent scarcity. Second, a hierarchical channel attention fusion neck (HCAF‐Neck) takes HPMCA‐enhanced features as input, integrating channel attention and adaptive fusion to effectively suppress background clutter, eliminate semantic discrepancies among multi‐level features, and prevent deep semantic information from overwhelming small‐object cues, thus improving fused feature quality. Third, a P2 detection branch is added to exploit high‐resolution spatial information from shallow maps, leveraging the optimized features from HCAF‐Neck to enhance perception and localization. Extensive experiments on VisDrone2019‐DET, TinyPerson, and SeaDronesSee show that our method achieves 37.5% mAP@50 on VisDrone, outperforming YOLOv12n by 6 percentage points, and also delivers competitive performance on the other datasets, confirming its stability and efficiency for small‐object detection.