基于改进的MobilenetV3热轧钢带表面缺陷分类  

Surface defect classification of hot rolled steel strips based on MobilenetV3 improvement

作　　者：熊政 车文刚[1] 保永莉 刘晓彤 XIONG Zheng;CHE Wen-gang;BAO Yong-li;LIU Xiao-tong(Key Laboratory of Computer Technology Application of Yunnan Province,Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Faculty of Life Science and Technology,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南省计算机技术应用重点实验室,昆明650500 [2]昆明理工大学生命科学与技术学院,昆明650500

出　　处：《兰州大学学报（自然科学版）》2024年第2期182-186,共5页Journal of Lanzhou University(Natural Sciences)

基　　金：云南省大学生创新创业训练计划项目(2021106740066)。

摘　　要：提出一种基于轻量化神经网络MobilenetV3-large改进的热轧钢带表面缺陷分类算法,通过剪枝、大量削减卷积层数、调整通道大小和步长,以及修改对应的网络参数快速降低了参数量.为弥补修改模型带来的准确率下降的问题,将激活函数ReLU更换为Hard-Swish,引入置换注意力机制替换原模型中的通道注意力机制,在进一步降低参数量的同时提高运行效率和分类准确率.在NEU-CLS表面缺陷数据集中的试验结果表明,改进后的算法参数量为0.5 MB,相比原模型降低96.89%,训练图片的时间由19.81 ms/幅降至10.73 ms/幅,平均准确率为99.26%,比改进前提高了5.56%,表明改进后的算法可应用于实时分类.An improved surface defect classification algorithm for hot-rolled steel strips was presented based on the lightweight neural network MobilenetV3-large.In order to quickly reduce the number of parameters,pruning was used,greatly reducing the number of convolution layers,adjusting the channel size and step size,and modifying the corresponding network parameters.In order to compensate for the decline in accuracy caused by the modification of the model,the activation function ReLU was modified to the Hard-Swish,and the shuffle attention mechanism was introduced to replace the squeeze-and-excitation attention mechanism in the original model to further reduce the number of parameters while improving the operating efficiency and classification accuracy.The experimental results showed that the parameters of the improved algorithm were 0.5 MB,96.89%less than the original model,the time spent training a picture reduced from 19.81 ms to 10.73 ms,and the average accuracy of the NEU-CLS surface defect dataset was 99.26%,being 5.56%higher than before the improvement and indicating that the improved algorithm can be applied to real-time classification.

关 键 词：MobilenetV3算法 转移注意力 结构性剪枝 缺陷分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]