A Multi-scale Network-based Method for the YOLOv3 Small Target Detection

Abstract

In order to further improve the accuracy of small target detection, this paper proposes a novel YOLOv3 small target detection method for multi-scale networks, which is mainly divided into four modules: 1. K-Means++ clustering algorithm to select anchor frames and accelerate model convergence; 2. multi- scale adaptive fusion to extract features and enhance network processing information; 3. end-to-end detection for network prediction to improve detection speed; 4. threshold score for ranking and using NMS to filter local maxima and output the predicted bounding box. Training and testing were conducted on the CCTSDB traffic sign dataset, and experiments showed that the algorithm significantly improved the detection accuracy of small targets compared with the original YOLOv3.