基于对抗性弱化的多阶段钢材表面缺陷分类算法
Adversarial Weakening Based Multi-Stage Classification for Steel Surface Defects
罗晶 1周威 1张昱中 1周雷1
作者信息
- 1. 上海理工大学健康科学与工程学院,上海 200093
- 折叠
摘要
针对钢材表面缺陷分类中缺陷特征类间差异小或类内差异大等难题,提出了一种多阶段钢材表面缺陷分类算法.首先,提出了一种基于图像拼接的数据预处理方法,在调整图像大小的同时保证了缺陷区域质量,增加了可识别的缺陷区域;然后,基于此预处理方法设计了一个多阶段分类网络来识别缺陷,网络中提出了一种基于对抗性弱化的缺陷区域挖掘模块,使得二阶段分类网络更侧重于一阶段分类网络未被注意的缺陷区域,从而计算得到更加完整的缺陷区域;最后,通过两阶段特征融合的方式得到网络的分类结果.此外,算法利用特征金字塔结构和通道-空间注意力机制设计了一种高效的特征提取网络用于各阶段分类网络实现高效的特征提取.在两个钢材缺陷数据集上进行了对比和消融实验,实验结果表明所提出的算法优于现有分类模型,同时在实际生产环境测试中也展示出良好的适用性.
Abstract
A multi-stage steel surface defect classification algorithm is proposed to address the challenges of small inter-class differences or large intra-class differences in defect features in steel surface defect classifi-cation.The paper first proposes a data preprocessing method based on image splicing,which ensures the quality of defective regions while resizing the image and increases the number of recognizable defective re-gions.A multi-stage classification network is then designed to identify defects based on this preprocessing method,and an adversarial weakening-based defect region mining module is proposed in the network,which makes the two-stage classification network focus more on the defective regions that have not been noticed by the one-stage classification network,so as to compute a more complete defective region.Finally,the classification results of the network are obtained by means of two-stage feature fusion.In addition the algo-rithm utilizes the feature pyramid structure and the channel-space attention mechanism to design an efficient feature extraction network for each stage of classification network to achieve efficient feature extraction.Comparison and ablation experiments are conducted on two steel defect datasets,and the experimental re-sults show that the proposed algorithm outperforms the existing classification models,as well as demon-strates good applicability in real production environment tests.
关键词
钢材缺陷分类/多阶段分类/对抗性弱化/特征金字塔/注意力机制Key words
classification of steel defects/multi-stage classification/weakness of antagonism/feature pyra-mid/attention mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据