MC-YOLOv5: A Multi-Class Small Object Detection Algorithm
Haonan Chen, Haiying Liu, Tao Sun, Haitong Lou, Xuehu Duan, Lingyun Bi, Lida Liu
Abstract
The detection of multi-class small objects poses a significant challenge in the field of computer vision. While the original YOLOv5 algorithm is more suited for detecting full-scale objects, it may not perform optimally for this specific task. To address this issue, we proposed MC-YOLOv5, an algorithm specifically designed for multi-class small object detection. Our approach incorporates three key innovations: (1) the application of an improved CB module during feature extraction to capture edge information that may be less apparent in small objects, thereby enhancing detection precision; (2) the introduction of a new shallow network optimization strategy (SNO) to expand the receptive field of convolutional layers and reduce missed detections in dense small object scenarios; and (3) the utilization of an anchor frame-based decoupled head to expedite training and improve overall efficiency. Extensive evaluations on VisDrone2019, Tinyperson, and RSOD datasets demonstrate the feasibility of MC-YOLOv5 in detecting multi-class small objects. Taking VisDrone2019 dataset as an example, our algorithm outperforms the original YOLOv5L with improvements observed across various metrics: mAP50 increased by 8.2%, mAP50-95 improved by 5.3%, F1 score increased by 7%, inference time accelerated by 1.8 ms, and computational requirements reduced by 35.3%. Similar performance gains were also achieved on other datasets. Overall, our findings validate MC-YOLOv5 as a viable solution for accurate multi-class small object detection.