基于YOLOv3模型的侧扫声呐沉船目标检测方法存在小目标漏警率高、模型权重大、检测速度未能满足实时性需求等问题,根据数据集特点,提出基于YOLOv5模型的侧扫声呐海底沉船目标检测方法.在YOLOv5模型的基础框架下,构建 YOLOv5a、YOLOv5b、YOLOv5c、YOLOv5d、YOLOv5s、YOLOv5m、YOLOv51 和 YOLOv5x8种不同深度和宽度的模型结构进行对比实验,并选择最优的结构,使用GA+K(genetic algorithm and K-means)算法优化检测框,并对损失函数进行改进.实验结果表明,改进的YOLOv5a模型在交并比阈值设置为0.5和0.5～0.95的平均准确率分别较原始模型提高了 0.3％和0.6％,较YOLOv3算法分别提高了 4.2％和6.1％,检测速度达到426帧/s,提升了近一倍,更加益于实际应用和工程部署.
An Improved YOLOv5 Method for Shipwreck Target Detection by Side-Scan Sonar Images
Objectives:The side-scan sonar shipwreck detection method based on the YOLOv3 model has the problems of high miss-alarm rate of small targets,heavy model weight,and slow detection speed that fails to meet real-time requirements.Methods:This paper introduces the YOLOv5 algorithm and pro-poses an improved YOLOv5 model according to the characteristics of the side-scan sonar shipwreck da-teset.We test YOLOv5a,YOLOv5b,YOLOv5c,YOLOv5d,YOLOv5s,YOLOv5m,YOLOv51 and YOLOv5x under the basic framework of YOLOv5 with eight different depth and width model structures.Then we choose the best structure by using genetic algorithm and K-means algorithm to optimize the detec-tion frame,and to improve the loss function through complete intersection over union.Results:The results show that under the different range of intersection over union as 0.5 and 0.5-0.95,the average precisions of the improved YOLOv5a model are increased by about 0.3％and 0.6％than that of the original model,re-spectively.Compared with the YOLOv3 model,the average precisions of the improved YOLOv5a model are increased by 4.2％and 6.1％,respectively,and the detection speed reaches 426 frames per second which is almost doubled that of YOLOv3.Conclusions:The proposed method is more conducive to practical appli-cations and engineering deployment.
side-scan sonarshipwreckobject detectionYOLOv5 modelloss functiongenetic algo-rithm and K-means algorithm
NETL
NSTL
万方数据
