基于轻量化深度学习VGG16网络模型的表面缺陷检测方法

Surface Defect Detection Method Based on Lightweight Deep Learning VGG16Net

零件表面缺陷是产品质量的重要组成部分,零件在线检测可以解决产品批量质量问题,已成为质量控制的未来发展趋势。人工检测方法存在检测精度低、漏检率高等问题,而通过机器视觉和深度学习相结合的在线检测成为研究热点。但是目前多数深度学习模型存在模型结构复杂、预测耗时长的缺点,不满足实时检测的需求。为了提高产品表面缺陷检测的效率和准确率,设计了一种轻量化深度学习VGG16网络模型,通过对VGG16Net进行结构优化与剪枝操作搭建轻量化识别模型,并构建表面缺陷数据集进行训练与预测。在标准表面缺陷数据集上,准确率达到0.949;搭建实验测试平台,对比原始VGG16网络,改进后的网络在剪枝率为50%时,准确率达到0.907,单张图片预测耗时为0.067s,模型压缩率为59.79%。轻量化表面缺陷检测方法具有高效自动、智能化等优点。

Surface defect is an important part of product quality.Online inspection can solve the problem of product quality,which has become the future trend of quality control.Manual detection methods have problems such as low detection accuracy and high missed detection rate.Online detection by machine vision and deep learning has become a research hotspot.However,most of the deep learning models have complex structures and long prediction time,which do not meet the requirement of real-time detection.In order to improve the efficiency and accuracy,this paper designs a lightweight deep learning VGG16Net by structural optimization and pruning operations,and constructs a surface defect dataset for training and prediction.On the standard surface defect dataset,the accuracy rate reaches 0.949.On the experimental test platform,with pruning rate of 50%,the final recognition accuracy rate reaches 0.907,single image prediction consumes 0.067 seconds,the compression rate is 59.79%.The lightweight surface defect detection method proposed in this paper has the advantages of high efficiency,automation and intelligence.