基于Transformer的陶瓷轴承表面缺陷检测方法  

Ceramic Bearing Surface Defect Detection Method Based on Transformer

作　　者：安冬[1] 胡荣华 王丽艳[1] 邵萌[1] 李新然 刘则通 AN Dong;HU Ronghua;WANG Liyan;SHAO Meng;LI Xinran;LIU Zetong(School of Mechanical Engineering,Shenyang Jianzhu University,Shenyang 110168,China)

机构地区：[1]沈阳建筑大学机械工程学院,沈阳110168

出　　处：《组合机床与自动化加工技术》2024年第2期160-163,168,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金面上项目(51975130);辽宁省教育厅项目(LJKMZ20220915)。

摘　　要：针对传统机器视觉检测方法中,由于陶瓷轴承滚动体表面曲率大、对比度低,表面成像模糊导致后续缺陷检测精度低的问题,提出一种基于Transformer的超分辨率残差网络。首先,网络使用残差学习策略,通过预测模糊图像与清晰图像之间的差值,实现超分辨率任务;其次,在网络上前端插入通道注意力模块和空间注意力模块并改进L2多头自注意力模块,以增强图像纹理、改善梯度爆炸问题;最后,针对超分辨率重建任务,提出一种两阶段训练策略优化训练过程。自建陶瓷轴承表面缺陷数据集上的大量实验结果表明,所提出网络模型在客观指标与主观评价上均优于MSESRGAN、VSDR等超分辨率算法,重建图像SSIM为0.939,PSNR为36.51 dB。A Transformer based super-resolution residual network is proposed for the problem of low accuracy of subsequent defect detection due to blurred surface imaging caused by large curvature and low contrast of ceramic bearing roller surface in traditional machine vision inspection methods.Firstly,the network uses a residual learning strategy to achieve the super-resolution task by predicting the difference between blurred and clear images;Secondly,a channel attention module and a spatial attention module are inserted in the front end of the network and the L2 multi-head self-attention module is improved to enhance the image texture and improve the gradient explosion problem;Finally,a two-stage training strategy is proposed to optimize the training process for the super-resolution reconstruction task.The extensive experimental results on the self-built ceramic bearing surface defect dataset show that the proposed network model outperforms super-resolution algorithms such as MSESRGAN and VSDR in both objective metrics and subjective evaluation,with a reconstructed image SSIM of 0.939 and PSNR of 36.51 dB.

关 键 词：Si_(3)N_(4)陶瓷轴承 超分辨率重建 TRANSFORMER 图像恢复 图像增强 

分 类 号：TH161[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]