基于小目标检测的YOLOv5的静止水面垃圾检测与分类算法  

Detection and Classification Algorithm of Static Water Garbage Based on YOLOv5 Small Target Detection

作　　者：李少杰 潘楚文 姚俊晖 贺景来 孙树平 

机构地区：[1]湖南理工学院信息科学与工程学院,湖南 岳阳

出　　处：《计算机科学与应用》2024年第4期219-229,共11页Computer Science and Application

摘　　要：针对现有水面垃圾检测模型精确率低、运算速度慢以及鲁棒性差的问题,提出一种基于改进型YOLOv5的水面垃圾检测与分类算法,将小目标检测头引入改进型网络框架以提高目标检测准确率,通过融入DeepSORT算法以提升检测速度,加入SENet注意力机制并更改损失函数为SIoU以提高模型的鲁棒性。通过对洞庭湖水域垃圾检测结果展开改进前后对比分析,实验结果表明,与原始YOLOv5相比,改进后的YOLOv5的mAP@0.5达到92.1%,较改进前提高了5.37%,同时平均帧率提高至47.8FPS,较改进前提升19.20%,为后续开展水面垃圾自适应处理可行性方案奠定了坚实基础。In response to the problems of low accuracy, slow operation speed, and poor robustness of existing water surface garbage detection models, a water surface garbage detection and classification algorithm based on improved YOLOv5 is proposed. The improved network framework introduces a small target detection head to improve target detection accuracy, incorporates the DeepSORT algorithm to enhance detection speed, adds SENet attention mechanism and changes the loss function to SIoU to improve the model's robustness. Through comparative analysis of the garbage detection results before and after improvement in the Dongting Lake area, the experimental results show that the improved YOLOv5 achieves an mAP@0.5 of 92.1%, an increase of 5.37% compared to before improvement, and the average frame rate is increased to 47.8 FPS, an increase of 19.20% compared to before improvement, laying a solid foundation for subsequent exploration of the feasibility of adaptive processing of water surface garbage.

关 键 词：YOLOv5 小目标检测头 网络框架 DeepSORT 注意力机制 损失函数 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]