基于Res-UNet算法的螺旋溜槽精矿带识别分割方法研究  

Research on Spiral Concentrator Concentrate Zone Identification Segmentation Method Based on Res-UNet

作　　者：刘惠中[1,2] 邓富龙 刘茜茜 刘建业 LIU Huizhong;DENG Fulong;LIU Xixi;LIU Jianye(School of Mechanical and Electrical Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi,China;Jiangxi Province Engineering Research Center for Mechanical and Electrical of Mining and Metallurgy,Ganzhou 341000,Jiangxi,China)

机构地区：[1]江西理工大学机电工程学院,江西赣州341000 [2]江西省矿冶机电工程技术研究中心,江西赣州341000

出　　处：《有色金属（选矿部分）》2024年第2期70-80,共11页Nonferrous Metals（Mineral Processing Section）

基　　金：国家自然科学基金资助项目(52164019);江西省“双千计划”引进高层次创新人才项目(jxsq2018101046)。

摘　　要：螺旋溜槽在铁、锡、钛、钽铌等金属及硫、煤等非金属矿的选矿生产中获大量应用,但目前螺旋溜槽的精矿截取调节控制还是依赖于人工,急需开发一种精矿的自适应截取技术代替人工截取以提高螺旋溜槽的生产效率。而实现这一目标的首要任务就是需要解决依赖人工肉眼获取精矿带位置信息的问题,因此提出了一个改进的UNet网络模型Res50-UNet-FD。算法模型使用UNet模型为基础,将残差网络ResNet50代替UNet网络中编码部分的特征提取网络,解决了深层特征提取过程中特征梯度消失以及网络消失的问题,有效提升了螺旋溜槽精矿带特征信息提取的精度。同时,为了改进和优化螺旋溜槽精矿带图像样本数据难易不平衡的问题,利用FocalLoss和DiceLoss的混合损失函数代替原本的CELoss损失函数。经对比,本文算法优于VGG-UNet、Res34-UNet、DC-UNet网络模型,算法模型的mIOU、mPA、F1分数和精确度分别为0.9632、0.9869、0.9870、0.9907。在性能指标上,本文算法无论是mIOU、mPA还是F1分数,整体性能都比VGG-UNet、Res34-UNet、DC-UNet网络模型高,算法的整体性能稳定。本文算法实现了对螺旋溜槽精矿矿带的分割识别,分割精度可以满足生产中对螺旋溜槽精矿分带特征信息识别的需求,为实现螺旋溜槽精矿的自适应截取奠定基础。Spiral concentrator has been widely applied in large scale in beneficiation process for iron,tin,titanium,tantalum niobium and other metals and sulfur,coal and other non-metallic ores in the beneficiation production has been a large number of applications,but at present the spiral concentrator concentrate adjustment relies on artificial control,there is an urgent need to develop an adaptive interception of concentrates instead of the artificial interception in order to improve the production efficiency of the spiral concentrator.The first task to achieve this goal is the need to solve the problem of relying on the artificial naked eye to obtain information about the location of the concentrate zone,so an improved UNet network model,Res50-UNet-FD,is proposed.The model uses the UNet model as the base model,and replaces the residual network ResNet50 with the feature extraction network in the coding part of the UNet network,which solves the problems of feature gradient disappearance as well as network disappearance in the process of deep feature extraction,and effectively improves the accuracy of feature information extraction in the concentrate zone of the spiral concentrator.The algorithm model uses the UNet model as the base model,and the residual network ResNet50 replaces the feature extraction network in the coding part of the UNet network,which solves the problems of feature gradient disappearance as well as network disappearance in the process of deep feature extraction,and effectively improves the accuracy of feature information extraction in the concentrate zone of the spiral concentrator.At the same time,in order to improve and optimize the problem of imbalance of the sample data of the spiral concentrator concentrate image,a hybrid loss function of Focal Loss and Dice Loss is utilized instead of the original Cross-Entropy loss function.Upon comparison,this paper's algorithm outperforms the VGG-UNet,Res34-UNet,and DC-UNet network models,and the mIOU,mPA,F1 score,and precision of the algorithm's model are 0.

关 键 词：螺旋溜槽 重力选矿 Res-UNet 矿带分割 

分 类 号：TD954[矿业工程—选矿] TD925.7