基于轻量型卷积视觉Transformer的锑浮选工况识别  

Working Condition Recognition Based on Lightweight Convolution Vision Transformer Network for Antimony Flotation Process

作　　者：陈奕霏 蔡耀仪[1] 李诗文 Chen Yifei;Cai Yaoyi;Li Shiwen(College of Engineering and Design,Hunan Normal University,Changsha 410083,Hunan,China)

机构地区：[1]湖南师范大学工程与设计学院,湖南长沙410083

出　　处：《激光与光电子学进展》2023年第6期249-261,共13页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61533021,61134006,61473319);湖南省自然科学基金(2020JJ5366)。

摘　　要：依靠人工观测锑浮选泡沫特征进行锑浮选工况识别,主观性强、误差大,严重制约浮选性能.基于计算机视觉的识别方法成本低、效果好.针对以上问题,提出一种基于轻量型卷积视觉Transformer(L-CVT)的锑浮选工况识别方法.通过Transformer层的堆叠代替标准卷积中矩阵乘法来学习全局信息,将卷积中的局部建模更替为全局建模,同时引入轻量型神经网络MobileNetv2中的子模块,减少计算成本.所提方法解决了卷积神经网络(CNN)忽略浮选图像内部长距离依赖关系的问题,同时也弥补了视觉Transformer(VIT)缺乏归纳偏置的缺点.实验结果表明,基于所提方法的锑浮选工况识别准确率最高可达93.56%,明显高于VGG16、ResNet18、AlexNet等主流网络,为锑浮选数据在工况识别领域提供了重要参考.It is highly subjective and has a large error to identify antimony flotation conditions by manually observing the characteristics of antimony flotation foam,which seriously restricts the flotation performance.The recognition method based on computer vision has low cost and good effect.In view of the above problems,a recognition method of antimony flotation conditions based on lightweight convolutional visual Transformer(LCVT)is proposed.The stack of transformer layers replaces matrix multiplication in standard convolution to learn global information,replaces local modeling in convolution with global modeling,and introduces submodules in the lightweight neural network MobileNetv2 to reduce computational costs.The proposed method solves the problem that convolutional neural network(CNN)ignores the longdistance dependence within flotation images,and makes up for the lack of inductive bias of visual Transformer(VIT).The experimental results show that the accuracy of antimony flotation condition identification based on the proposed method can reach 93.56%,which is significantly higher than VGG16,ResNet18,AlexNet and other mainstream networks.It provides an important reference for antimony flotation data in the field of condition identification.

关 键 词：机器视觉 锑浮选 工况识别 计算机视觉 轻量型卷积神经网络 视觉Transformer 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]