基于LIR和GFNet的带钢表面缺陷识别
Strip Steel Surface Defect Recognition Based on LIR and GFNet
刘双辉 1易灿灿 1肖涵 1黄涛1
作者信息
- 1. 武汉科技大学 冶金装备及其控制教育部重点实验室,武汉 430081;武汉科技大学 机械传动与制造工程湖北省重点实验室,武汉 430081;武汉科技大学 精密制造研究院,武汉 430081
- 折叠
摘要
针对深度学习(deep learning,DL)模型处理带钢表面缺陷图像存在计算成本大、实时性差的问题,提出了一种基于可学习的图像调整器(learnable image resizer,LIR)和扫视-聚焦网络(glance and focus network,GFNet)的带钢表面缺陷分类方法.首先,针对DL模型在处理带钢表面缺陷图像时存在空间冗余的问题,提出GFNet驱动的带钢表面缺陷识别模型,其可以根据不同样本自适应分配计算资源,在模型推理阶段显著减少计算量;其次,提出LIR和GFNet联合训练的方法,调整图像大小的同时实现针对识别模型的特征增强;最后,收集整理了某钢铁企业冷轧薄板厂带钢表面缺陷数据集,利用所提方法进行分析.将残差网络(residual networks,ResNet)的ResNet-50 模型作为主干网络,与原始ResNet-50 比较,所提方法在不牺牲准确率的情况下,将单张图像的推断时间减少约3.58 倍,计算量降低约6.11 倍,从而验证了提出方法的有效性.
Abstract
To address the problems of large computational cost and poor real-time performance of deep learning(DL)models in strip surface defect image recognition,a strip surface defect recognition method based on learnable image resizer(LIR)and glance and focus network(GFNet)is proposed.First,to ad-dress the problem of spatial redundancy of DL model in processing strip steel surface defect images,a GF-Net-driven strip steel surface defect recognition model is proposed,which can adaptively allocate computa-tional resources according to different samples and significantly reduce the computational effort in the mod-el inference stage;then,a joint training method of LIR and GFNet is proposed to realize feature enhance-ment for the recognition model while adjusting the image size Finally,a dataset of strip steel surface defects in a cold rolled sheet mill of an iron and steel enterprise is collected and analyzed using the proposed meth-od.Using the ResNet-50 model of residual networks(ResNet)as the backbone network and comparing it with the original ResNet-50,the proposed method reduces the inference time of a single image by about 3.58 times and the computational effort by about 6.11 times without sacrificing the accuracy,thus valida-ting the effectiveness of the proposed method.
关键词
带钢表面缺陷/图像分类/可学习的图像调整器/动态神经网络/扫视-聚焦网络Key words
strip steel surface defects/image classification/learnable image resizer(LIR)/dynamic neural networks/glance and focus network(GFNet)引用本文复制引用
基金项目
国家自然科学基金项目(51805382)
湖北省重点研发计划项目(2021BAA194)
出版年
2024
NETL
NSTL
维普
万方数据