基于轻量化二值神经网络的钢铁表面缺陷分类  

Steel Surface Defect Classification Based on Lightweight Binarized Neural Networks

作　　者：卓晨涛 吴丽君[1] ZHUO Chentao;WU Lijun(College of Physics and Information Engineering,Fuzhou University,Fuzhou350108,CHN)

机构地区：[1]福州大学物理与信息工程学院,福州350108

出　　处：《光电子技术》2024年第4期317-323,共7页Optoelectronic Technology

基　　金：国家自然科学基金项目(62271151);福建省自然科学基金项目(2021J01580)。

摘　　要：基于二值化网络提出了轻量二值化钢铁缺陷分类网络(Lightweight Binarized Steel Defect Classification Network,LBSDC-Net),以期实现实时高精度的钢铁缺陷自动分类。首先,基于可变阈值符号函数和组卷积的理念,设计了双阈值型组卷积模块,以在压缩网络模型的同时最小化二值组卷积引起的信息损失,将基础网络模型大小降低了31.2%,精度仅下降0.34%;其次,通过调整下采样卷积的步长并结合最大池化,降低了残差网络中捷径分支下采样时的信息损失,提升了网络的分类性能;在NEU-CLS钢铁缺陷数据集上的实验结果表明,网络模型大小为11.86 MBit时,LBSDC-Net网络在钢铁缺陷分类任务中准确率达到99.06%。相较于基础网络Bi-Real-Net98.73%的准确率和17.23 MBit的网络模型大小,LBSDC-Net实现了网络规模的有效压缩,还提升了分类精度。A Lightweight Binarized Steel Defect Classification Network(LBSDC-Net)was pro-posed to achieve real-time high-precision automatic classification of steel defects.Firstly,a dual-threshold grouped convolution module was designed based on the concepts of variable threshold sign function and group convolution.This design minimized the information loss caused by binarized group convolution while compressing the network model,reducing the base network model size by 31.2%with only a 0.34%decrease in accuracy.Secondly,the classification performance of the network was improved by adjusting the stride of the down-sampling convolution and combining it with max pooling to reduce information loss during down-sampling in the shortcut branch of the residual network.Exper-imental results on the NEU-CLS steel defect dataset showed that with a model size of 11.86 MBit,the LBSDC-Net achieved an accuracy of 99.06%in the steel defect classification task.Compared to the base network Bi-Real-Net,which had an accuracy of 98.73%and a model size of 17.23 MBit,the LBSDC-Net effectively compressed the network scale while also improving classification accuracy.

关 键 词：二值神经网络 轻量化 深度学习 钢铁缺陷 分类 

分 类 号：TN911.7[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]