基于多样性激励GAN及MobileViT的带钢缺陷分类方法  

Strip Steel Defect Classification Method Based on Variety-Stimulated Generative Adversarial Network and MobileViT

作　　者：刘劲豪 金红[1,2] 谢晓兰[1,2] 何礼安 LIU Jinhao;JIN Hong;XIE Xiaolan;HE Li′an a(College of Computer Science and Engineering,Guilin University of Technolog,Guilin 541006,china;Guangxi Key Laboratory of Embedded Technology and Intelligent System,Guilin University of Technolog,Guilin 541006,china)

机构地区：[1]桂林理工大学计算机科学与工程学院,桂林541006 [2]桂林理工大学广西嵌入式技术与智能系统重点实验室,桂林541006

出　　处：《组合机床与自动化加工技术》2025年第4期183-187,194,共6页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(62262011);广西重点研发计划项目(桂科AB23049001)。

摘　　要：针对在工业环境下因图片数据样本量较小影响带钢表面缺陷检测分类精度,通常采用数据增广策略但需保证细节特征多样性的问题,提出了一种多样性激励深度卷积生成对抗网络VS-DCGAN(variety-stimulation deep convolutional generation adversal network)和MobileViT的带钢表面缺陷分类方法。首先,引入深度卷积生成对抗网络,利用通道注意力机制和多样性激励模块对其进行改进;再用VS-DCGAN对原始数据集进行扩充并保证细节特征多样性;然后,在扩充后的带钢表面缺陷数据集上训练轻量级分类网络MobileViT;最终实现带钢表面缺陷的分类。实验结果表明,多样性激励深度卷积生成对抗网络可以生成与原始图像数据分布相似且细节特征更丰富的带钢表面缺陷图片,达到了数据增强的目的;且扩充后的数据集能够使得轻量级分类网络MobileViT的训练效果更优异,3个版本的分类准确率分别为97.33%、98.66%、97.33%,分别提高了0.33%、1.5%、2.13%,高效地实现了带钢表面缺陷分类。To address the issue of low sample size affecting the accuracy of steel strip surface defect detection and classification in industrial environments,a method for steel strip surface defect classification based on the variety-stimulation deep convolutional generative adversarial network(VS-DCGAN)and MobileViT is proposed.This method aims to enhance the data while ensuring the diversity of detailed features.First,a deep convolutional generative adversarial network is introduced and improved using channel attention mechanisms and variety-stimulation modules.Then,VS-DCGAN is used to augment the original dataset while preserving the diversity of detailed features.Subsequently,the lightweight classification network MobileViT is trained on the augmented steel strip surface defect dataset to achieve defect classification.Experimental results show that the VS-DCGAN can generate steel strip surface defect images that are similar to the original image data distribution and richer in detailed features,achieving the goal of data augmentation.Additionally,the augmented dataset enables the lightweight classification network MobileViT to achieve superior training results,with classification accuracies of 97.33%,98.66%,and 97.33%for the three versions,respectively,representing improvements of 0.33%,1.5%,and 2.13%,effectively realizing steel strip surface defect classification.

关 键 词：图像处理 图像分类 生成对抗网络 数据增强 

分 类 号：TH165[机械工程—机械制造及自动化] TG66[金属学及工艺—金属切削加工及机床]