基于改进YOLOv5的矿石目标检测算法

Ore Target Detection Algorithm Based on Improved YOLOv5

摘要: 针对工业矿山生产线上工作量大，在进行目标定位检测时需兼具较高实时性，提出一种基于改进YOLOv5的目标检测算法。该算法在主干特征提取网络引入更轻量化的MobileNetv3，同时将PANet网络3 × 3的卷积块改进，提高了检测精度和效率；使用DIoU-NMS替代原模型中NMS的方式，提高矿石重叠时检测精度。改进后的算法的参数量减少了24.95%，mAP值达到了95.7%，帧数达到了166.7 FPS，检测单张图片的时间缩短了4 ms，表明该算法在矿石传送带上具有更强实用性和更高的检测精度。

Abstract: Aiming at the heavy workload in the industrial mine production line and the high real-time performance in the target positioning detection, a target detection algorithm based on improved YOLOv5 is proposed. The algorithm replaces the backbone feature extraction network CSPDarknet with the lighter MobileNetv3, and improves the 3 × 3 convolutional block of PANet network to improve the detection accuracy and efficiency. Using DIoU-NMS to replace the weighted NMS in the original model further reduces the number of parameters, improves the convergence of the model, and improves the detection accuracy when the targets overlap. The parameter number of the improved algorithm is reduced by 24.95%, the mAP value is 95.7%, the frame number is 166.7 FPS, and the detection time of a single image is shortened by 4 ms, which shows that the algorithm has stronger practicability and higher detection accuracy on the ore conveyor belt.