基于目标图像组合算法的改进YOLOv5模型  

An improved YOLOv5 model based on target image fusion algorithm

作　　者：张艳伟[1,2] 周朝广 黄一帆 曹菁菁 ZHANG Yanwei;ZHOU Chaoguang;HUANG Yifan;CAO Jingjing(School of Transportation and Logistics Engineering,Wuhan University of Technology,Wuhan 430063,Hubei,China;Engineering Research Center of Port Logistics Technology and Equipment,Ministry of Education,Wuhan 430063,Hubei,China)

机构地区：[1]武汉理工大学交通与物流工程学院,湖北武汉430063 [2]港口物流技术与装备教育部工程研究中心,湖北武汉430063

出　　处：《中国工程机械学报》2023年第6期509-514,共6页Chinese Journal of Construction Machinery

基　　金：科技创新2030重大项目(2022ZD0119303)。

摘　　要：针对YOLOv5目标检测模型训练时间长、检测精度偏低问题,提出一种目标图像组合算法,考虑必要的图像背景及图像覆盖对目标图像进行分割,设计减少图像失真的重组策略提高单张训练集图像内目标个数,降低模型训练时长。改进先验框生成策略,以绝对差值作为距离函数,对训练集目标边框的长和宽分别进行一维K-means聚类,提高先验框对训练集的适应度。提出多层并列卷积结构,对输入特征经过三层并列卷积后的输出进行融合,增强特征表征能力。以VOC2007和VOC2012训练集和验证集作为训练图像,采用目标图像组合算法,模型训练时间减少30%以上,改进先验框生成策略使先验框对训练集的适应度达到0.735。在VOC2007测试数据集上测试,改进YOLOv5模型平均准确率均值(mAP)由79.1%提升至80.3%。Aiming at the problem of low detection precision and long training time of YOLOv5,a target image fusion algorithm is proposed.Considering the necessary background images and image coverage,target images are segmented,and the target image recombination strategy with reducing image distortion is designed to improve the number of targets within a single training image and decrease the model training time.An anchor boxes generation strategy is improved with absolute distance being used as a distance function for one-dimensional K-means clustering of targets’length and width respectively to improve anchor boxes’fitness to the training set.The results generated by three parallel convolution layers are fused in multi-parallel convolutional network(MPCNet)to enhance the ability of feature characterization.VOC2007 and VOC2012 training and validation sets are used as a training set,the model training time is reduced more than 30%by using the target image fusion algorithm.With the improved anchor boxes generation strategy,fitness of anchor boxes to the training set is improved to 0.735.Tested on the VOC2007 datasets,the mean average precision of the improved YOLOv5 is increased from 79.1%to 80.3%.

关 键 词：目标检测 YOLOv5 图像分割 多层并列卷积 先验框生成 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]