引用本文:王子轩,汤健,夏恒,张晓晓,荆中岭,韩红桂.基于并行差分进化–梯度特征深度森林的废旧手机识别方法[J].控制理论与应用,2022,39(11):2137~2148.[点击复制]
WANG Zixuan,TANG Jian,XIA Heng,ZHANG Xiaoxiao,JING ZHongling,HAN Honggui.Used mobile phone recognition method based on parallel differential evolution and gradient feature deep forest[J].Control Theory and Technology,2022,39(11):2137~2148.[点击复制]
基于并行差分进化–梯度特征深度森林的废旧手机识别方法
Used mobile phone recognition method based on parallel differential evolution and gradient feature deep forest
摘要点击 1130  全文点击 341  投稿时间：2021-07-29  修订日期：2022-08-09
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2022.10687
  2022,39(11):2137-2148
中文关键词  手机回收装备  废旧手机识别  并行差分进化  深度森林  深度学习
英文关键词  mobile phone recycling equipment  used mobile phone (UMP) recognition  parallel differential evolution (PDE)  deep forest (DF)  deep learning (DL)
基金项目  国家重点研发计划项目(2018YFC1900800–5), 国家自然科学基金项目(62073006, 61573364, 61873009)资助.
作者单位邮编
王子轩 北京工业大学信息学部 100024
汤健* 北京工业大学信息学部 100124
夏恒 北京工业大学信息学部 
张晓晓 北京抱扑再生环保科技有限公司 
荆中岭 北京抱扑再生环保科技有限公司 
韩红桂 北京工业大学 信息学部 
中文摘要
      废旧电子产品“互联网+回收”模式的推广, 使得无人化、智能化的废旧手机(UMP)回收装备成为典型城市 固体废物资源化领域的重点关注对象. 本文以基于回收装备的UMP智能化识别组件为研究对象, 设计并实现了一 种基于并行差分进化(PDE)–梯度特征深度森林(GfDF)算法的UMP识别方法. 本方法由UMP识别模型和PDE参数寻 优模型组成, 其中: 前者包含的UMP定位裁剪模块基于Faster–RCNN模型对图像裁剪以获得有效信息, GfDF识别模 块通过引入多尺度梯度特征策略使其更易学习“定位模块”抓取信息; 后者使用并行策略优化GfDF模型超参数以 提高UMP识别精度. 实验结果表明, 相比于深度模型和其他机器学习模型, 本方法在识别精度和训练时间上均具有 优势, 能够有效提高回收装备自动化程度和手机回收效率.
英文摘要
      The promotion of the “Internet + Recycling” model of waste electronic products has made unmanned and intelligent used mobile phone (UMP) recycling equipment become the focus of attention in the field of typical urban solid waste recycling. This papper takes used mobile phone recognition (UMPR) based on recycling equipment as the research object. We design and implement a UMPR method based on parallel differential evolution (PDE)-gradient feature deep forest (GfDF) algorithm. This method is composed of the UMPR model and the PDE parameter optimization model. The mobile phone positioning and cropping module included in the former is based on the Faster–RCNN model and crops the image to obtain the effective information. The GfDF recognition module introduces a multi-scale gradient feature strategy to make it easier to learn the “location module” for capturing information. The PDE parameter optimization module uses a parallel strategy to optimize the hyperparameters of the GfDF model to improve the accuracy of the UMP identification. The experimental results show that compared with deep models and other machine learning models, this method has performance advantages in recognition accuracy and training time. It can effectively improve the degree of automation of recycling equipment and the efficiency of mobile phone recycling.