An Improved YOLOv3-SPP Algorithm for Image-Based Pothole Detection

This paper proposes an improved YOLOv3-SPP crater target detection algorithm to address the issues of complex and variable backgrounds, diverse targets, and difficulty in detecting small targets in military scene object detection tasks. By adding a prediction layer (Yolo 4) to the YOLOv3-SPP network structure, the detection performance for small targets can be improved. In order to extract richer semantic information of the network, this paper replaces the pooling structure in the SPP module with softpool. Then, this paper introduces CIoU bounding box regression loss to improve positioning accuracy and uses the K-Means++ clustering algorithm to generate the optimal anchor box. Finally, the proposed method is verified through a publicly available pothole datasets to illustrate the performance of the proposed algorithm.