Journal of Computer Applications ›› 2021, Vol. 41 ›› Issue (6): 1836-1841.DOI: 10.11772/j.issn.1001-9081.2020081244

Special Issue: 前沿与综合应用

• Frontier and comprehensive applications • Previous Articles     Next Articles

Classification of steel surface defects based on lightweight network

SHI Yangxiao1, ZHANG Jun1, CHEN Peng1, WANG Bing2   

  1. 1. School of Electrical Engineering and Automation, Anhui University, Hefei Anhui 230601, China;
    2. School of Electrical and Information, Anhui University of Technology, Ma'anshan Anhui 243002, China
  • Received:2020-08-18 Revised:2020-10-27 Online:2021-06-10 Published:2020-12-09
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61872004).

基于轻量级网络的钢铁表面缺陷分类

史杨潇1, 章军1, 陈鹏1, 王兵2   

  1. 1. 安徽大学 电气工程与自动化学院, 合肥 230601;
    2. 安徽工业大学 电气信息学院, 安徽 马鞍山 243002
  • 通讯作者: 史杨潇
  • 作者简介:史杨潇(1997-),男,安徽合肥人,硕士研究生,主要研究方向:深度学习、缺陷检测;章军(1971-),男,安徽合肥人,教授,博士,主要研究方向:深度学习在视频分析中的应用、化学信息学;陈鹏(1978-),男,安徽肥东人,教授,博士,主要研究方向:机器学习、生物信息学;王兵(1976-),男,安徽怀宁人,副教授,博士,主要研究方向:模式识别、生物信息学。
  • 基金资助:
    国家自然科学基金资助项目(61872004)。

Abstract: Defect classification is an important part of steel surface defect detection. When the Convolutional Neural Network (CNN) has achieved good results, the increasing number of network parameters consumes a lot of computing cost, which brings great challenges to the deployment of defect classification tasks on personal computers or low computing power devices. Focusing on the above problem, a novel lightweight network model named Mix-Fusion was proposed. Firstly, two operations of group convolution and channel-shuffle were used to reduce the computational cost while maintaining the accuracy. Secondly, a narrow feature mapping was used to fuse and encode the information between the groups, and the generated features were combined with the original network, so as to effectively solve the problem that "sparse connection" convolution hindered the information exchange between the groups. Finally, a new type of Mixed depthwise Convolution (MixConv) was used to replace the traditional DepthWise Convolution (DWConv) to further improve the performance of the model. Experimental results on NEU-CLS dataset show that, the number of floating-point operations and classification accuracy of Mix-Fusion network in defect classification task is 43.4 Million FLoating-point Operations Per second (MFLOPs) and 98.61% respectively. Compared to the networks of ShuffleNetV2 and MobileNetV2, the proposed Mix-Fusion network reduces the model parameters and compresses the model size effectively, as well as obtains the better classification accuracy.

Key words: surface defect detection, defect classification, model acceleration, deep learning, lightweight network

摘要: 缺陷分类是钢铁表面缺陷检测的重要内容。在卷积神经网络(CNN)取得良好效果的同时,网络日益增长的参数量耗费了大量计算成本,为缺陷分类任务在个人计算机或低算力设备上的部署带来了巨大的挑战。针对上述问题,提出了一种新颖的轻量级网络模型Mix-Fusion。首先,通过组卷积和通道洗牌两种操作,在保持精度的同时有效降低计算成本;其次,利用一个狭窄的特征映射对组间信息进行融合编码,并将生成的特征与原始网络结合,从而有效解决了“稀疏连接”卷积阻碍组间信息交换的问题;最后,用一种新型的混合卷积(MixConv)替代了传统的深度卷积(DWConv),以进一步提高模型的性能。在NEU-CLS数据集上的实验结果表明,Mix-Fusion网络在缺陷分类任务中的浮点运算次数和分类准确率分别为43.4 MFLOPs和98.61%。相较于ShuffleNetV2和MobileNetV2网络,Mix-Fusion网络不仅降低了模型参数,压缩了模型大小,同时还得到了更好的分类精度。

关键词: 表面缺陷检测, 缺陷分类, 模型加速, 深度学习, 轻量级网络

CLC Number: