An Improved YOLOv5 Multi-Scale Object Detection Method
Xin Liu, Yihan Wang, Xue Chen, Feng Chen · 2023
Multi-scale object detection often faces challenges with lower accuracy. We have created a solution for multi- scale object detection named MS-YOLOv5s. The approach focuses on enhancing multi-scale feature extraction, rectifying the absence of multi-scale features in feature extraction, and addressing deficiencies in feature fusion, particularly for exploiting low-level features. First, the Multi-path_C3 multi-branch feature extraction structure is constructed to replace the C3 structure in YOLOv5s. This multi-branch feature extraction structure is employed to capture multi-scale features, and the captured multi-scale features are utilized to supplement contextual information, thereby improving the effectiveness of object detection. Secondly, a feature fusion structure is constructed to incorporate low-level features. In this paper, the fusion paths within the YOLOv5s feature fusion structure are enhanced to alleviate the loss of low-level features resulting from frequent sampling. In this paper, we experimentally showcase the strong performance of MS-YOLOv5s in multi- scale object detection across various datasets.