China Mechanical Engineering ›› 2022, Vol. 33 ›› Issue (19): 2298-2305,2316.DOI: 10.3969/j.issn.1004-132X.2022.19.003

Previous Articles     Next Articles

Metal Defect Image Recognition Method Based on Shallow CNN Fusion Transformer

TANG Donglin;YANG Zhou;CHENG Heng;LIU Mingxuan;ZHOU Li;DING Chao   

  1. School of Mechanical and Electrical Engineering,Southwest Petroleum University,Chengdu,610500
  • Online:2022-10-10 Published:2022-10-20

浅层卷积神经网络融合Transformer的金属缺陷图像识别方法

唐东林;杨洲;程衡;刘铭璇;周立;丁超   

  1. 西南石油大学机电工程学院,成都,610500
  • 作者简介:唐东林,男,1970年生,教授,博士后研究人员。研究方向为图像处理和模式识别。E-mail:tdl840451816@163.com。
  • 基金资助:
    中国职业安全健康协会创新创业项目(CXCY-2021-22);四川省科技支撑计划(2017FZ0033);成都市技术创新研发项目(2018-YF05-00201-GX);西南石油大学国家重点实验室项目(PLN201828)

Abstract: Aiming at the problems of large amount of parameters and calculation in the field of metal defect recognition, a metal defect recognition method was proposed based on shallow CNN and Transformer model. The shallow CNN was used to learn the local information and position information of the images, and the Transformer was used to learn the global information of the images. At the same time, the channel attention module SE was introduced to pay attention to the important feature channels to realize the defect image recognition. The effectiveness of this method was verified by introducing the open defect data set, and the universality of this method was verified by using the self-built defect ultrasonic data set. The experimental results show that the proposed method has strong universality and may effectively identify metal defect images on small and medium-sized data sets. 

Key words: metal defect identification, deep learning, convolutional neural network(CNN), Transformer model, multi-head attention

摘要: 针对金属缺陷识别领域中传统深度学习方法存在参数量多、计算量大的问题,提出了一种浅层卷积神经网络融合Transformer模型的金属缺陷识别方法。利用浅层卷积神经网络学习图像局部信息与位置信息,通过Transformer学习图像全局信息,同时引入通道注意力模块SE关注重要特征通道,实现缺陷图像识别。通过引入公开缺陷数据集验证该方法的有效性,同时利用自建缺陷超声数据集验证所提方法的通用性。实验结果表明,在中小规模数据集上,该方法通用性较强,能够对金属缺陷图像进行有效识别。

关键词: 金属缺陷识别, 深度学习, 卷积神经网络, Transformer模型, 多头注意力

CLC Number: