李成,汪小凯,章振原,秦训鹏.基于深度残差网络的有色金属破碎料视觉识别方法[J].有色金属工程,2019,9(8):.
基于深度残差网络的有色金属破碎料视觉识别方法
A method of visual recognition for nonferrous metal crushed aggregates based on deep residual networks
投稿时间:2019-01-12  修订日期:2019-01-22
DOI:
中文关键词:  深度学习  视觉识别  有色金属破碎料
英文关键词:deep learning  visual recognition  nonferrous metal crushed aggregates
基金项目:湖北省重大科技创新计划项目(2015AAA014)
           
作者单位
李成 武汉理工大学
汪小凯 武汉理工大学
章振原 武汉理工大学
秦训鹏 武汉理工大学
摘要点击次数: 39
全文下载次数: 42
中文摘要:
      从报废机械中回收和分选有色金属对于环保和资源节约具有重大意义。本文提出一种基于深度残差网络的有色金属破碎料视觉识别方法,使分选过程自动化。为了从彩色CCD相机采集的图像中获得金属破碎料的感兴趣区域(ROI),本文通过反距离加权插值算法优化大津方法,使得图像分割的阈值可以自适应的调节,以取得更好的分割效果。有色金属破碎料的感兴趣区域作为一个39层的深度残差网络的输入用以训练网络。深度残差网络输出每个有色金属破碎料的类别。 工控机通过这种算法和相机标定获取破碎料在传送带上的位置和类别信息来控制分选机构。作为对比,本文开展基于传统的卷积网络,如AlexNet,VGGNet16,VGGNet39,GoogLeNet (Inception V2)的有色金属破碎料识别测试。实验表明,基于深度残差网络的方法对铝料、铜料和其他杂料的识别准确率分别达到98.7%、98.9%、96.2%,优于基于传统的卷积网络。
英文摘要:
      Recycling and sorting nonferrous metals from scrap machinery is of great significance for environmental protection and resource conservation. This paper proposes a visual recognition method for nonferrous metal crushing materials based on deep residual networks, so as to make the sorting process automatic. In order to obtain the regions of interest (ROIs) of the metal crushing materials from the images acquired by the color CCD camera, Otsu method is improved through inverse-distance weighting interpolation algorithm, so that the threshold of image segmentation can be adaptively adjusted for better segmenting result. The ROIs of nonferrous metal crushed aggregates as the input signals of a 39-layer deep residual networks are used to train the networks. The deep residual networks output the categories of each nonferrous metal crushed aggregate. The industrial computer controls the sorting mechanism according to the positions and categories of each nonferrous metal crushed aggregate on the conveyor belt through this algorithm and camera calibration. For comparison, this paper carried out the recognition test of nonferrous metal crushed aggregate based on the traditional convolutional networks, such as AlexNet, VGGNet16, VGGNet39, GoogLeNet (Inception V2). Experiments showed that the recognition accuracies of the method based on deep residual networks for aluminum, copper and other materials are respectively 98.7%, 98.9% and 96.2%, which are superior to those based on the traditional convolutional networks.
查看全文  查看/发表评论  下载PDF阅读器
关闭