废钢料型智能识别的语义分割模型选择与实现  

Selection and implementation of semantic segmentation model for intelligent recognition of scrap type

作　　者：朱立光 陈泊羽 肖鹏程 张妍 许云峰[3] ZHU Liguang;CHEN Boyu;XIAO Pengcheng;ZHANG Yan;XU Yunfeng(College of Metallurgy and Energy,North China University of Science and Technology,Tangshan 063210,China;School of Material Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050021,China;School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050021,China)

机构地区：[1]华北理工大学冶金与能源学院,河北唐山063210 [2]河北科技大学材料科学与工程学院,河北石家庄050021 [3]河北科技大学信息科学与工程学院,河北石家庄050021

出　　处：《冶金自动化》2023年第3期81-92,共12页Metallurgical Industry Automation

基　　金：国家自然科学基金青年科学基金项目(51904107);河北省自然科学基金优秀青年科学基金项目(E2020209005,E2021209094);河北省高等学校科学技术研究项目(BJ2019041);河北省“三三三人才工程”资助项目(A202102002);唐山市人才资助重点项目(A202010004)。

摘　　要：目前业界对废钢智能评级和扣杂的方案都是基于目标识别模型,但是目标识别模型与语义分割模型相比其无法精确刻画废钢的边界,造成对废钢的面积估计和特征采集不精准。然而当前语义分割模型众多,如何选择一种适合废钢判级场景的模型是亟待解决的问题。针对该问题,首先在实验室建立1∶3物理模型,模拟不同类型废钢入厂验质场景;然后,用2K分辨率摄像头采集图像数据;最后,将主流的20种语义分割模型进行对比分析。试验表明,在139张废钢数据集上应使用全卷积神经网络(fully convolutional network,FCN)模型搭配高分辨率网络(high-resolution net,HRNet)对废钢进行语义分割;在图像增强的1529张废钢数据集上应使用基于Transformer改进的高效语义分割模型SegFormer-B5对废钢进行预测分类;从平均交并比(mean intersection over union,mIoU)评价指标来看,FCN和基于上下文表示的语义分割网络(object-contextual representations network,OCRNet)使用HRNet主干网络比残差主干网络(residual network,ResNet)平均提高6.6个百分点。At present,the intelligent rating and impurity deduction schemes for scrap in the industry are based on the target recognition model,but compared with the semantic segmentation model,the target recognition model cannot accurately depict the boundary of scrap,resulting in inaccurate area estimation and feature collection of scrap.However,there are many semantic segmentation models at present.How to select a model suitable for scrap grading scenarios is a problem to be solved.To solve this problem,a 1∶3 physical model was established in the laboratory to simulate different types of scrap entering the factory for quality inspection.Then 2K resolution camera was used to collect image data.Finally,20 mainstream semantic segmentation models were compared and analyzed.The experiment shows that on 139 scrap datasets,full convolutional network(FCN)model and high resolution net(HRNet)are used to segment scrap semantically.The improved efficient semantic segmentation model SegFormer-B5 based on Transformer should be used on the 1529 scrap steel data sets with image enhancement to predict and classify scrap.From the mean intersection over union(mIoU)index,the HRNet backbone network used by FCN and object-contextual representations network(OCRNet)is 6.6 percentage points higher than the residual network(ResNet).

关 键 词：废钢 语义分割 FCN OCRNet SegFormer-B5 Deeplabv3P 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TF702.3[自动化与计算机技术—计算机科学与技术]