基于集成自监督的工业视觉大模型算法研究  

Research on Large Industrial Vision Models Based on Ensemble Self-supervised Learning

作　　者：林而贤 张潮 周雄图[1] 张永爱[1] LIN Erxian;ZHANG Chao;ZHOU Xiongtu;ZHANG Yongai(College of Physics&Information Engineering,Fuzhou University,Fuzhou 350116,China;BYD Automotive Industry Limited Company,Shenzhen 518118,China)

机构地区：[1]福州大学物理与信息工程学院,福州350116 [2]比亚迪汽车工业有限公司,广东深圳518118

出　　处：《小型微型计算机系统》2025年第4期907-913,共7页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2021YFB3600603)资助;福建省自然科学基金项目(2020J01468)资助;国家自然科学基金青年科学基金项目(62101132)资助.

摘　　要：针对工业视觉检测场景多样,模型迁移相较其他计算机视觉任务频繁更多,同时工业质检领域还存在小样本,缺乏异常数据的问题,本文将掩码自编码器(Masked Autoencoders,MAE)与对比自监督相结合,提出了一种适用于工业大模型训练的集成自监督算法(Integrated Autoencoders,IAE),在工业场景下开发原创的工业视觉大模型训练方法以帮助工业视觉大模型在工业图片上抽取更好的特征,提高下游任务图像分类准确率,以及算法训练效率和模型的泛化性.对比传统MAE算法,本文提出的IAE算法训练出来的大模型预训练权重在迁移到同一下游工业站点的小数据集分类任务上精度提高了2.42%;大数据集分类任务上精度提高了0.86%.In the context of diverse industrial visual inspection scenarios,where model transfer is more frequent compared to other computer vision tasks,and the industrial quality inspection domain faces challenges such as small sample sizes and a lack of anomaly data,this paper combines Masked Autoencoders(MAE)with contrastive self-supervision to propose an integrated self-supervised algorithm called Integrated Autoencoders(IAE).This approach is developed for large-scale industrial model training,aiming to help these models extract superior features from industrial images.The objective is to enhance downstream tasks such as image classification accuracy,algorithm training efficiency,and model generalization in the industrial setting.In comparison to traditional Mask Autoencoder(MAE)algorithm,the large models pretrained with the IAE algorithm exhibited a 2.42%increase in accuracy when transferred to small dataset classification tasks at the same downstream industrial site.Furthermore,they demonstrated a 0.86%accuracy improvement on large dataset classification tasks.

关 键 词：工业视觉 IAE 大模型 预训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]