一种基于CBAL2-Net卷积网络的轻量级X射线铜矿石分选方法  

A Lightweight X-ray Copper Ore Sorting Method Based on CBAL2-Net Convolutional Network

作　　者：孔钰文 何剑锋[1,2] 朱文松 李卫东[1,2] 王杉 汪雪元[1,2] 钟国韵 瞿金辉[1] KONG Yuwen;HE Jianfeng;ZHU Wensong;LI Weidong;WANG Shan;WANG Xueyuan;ZHONG Guoyun;QU Jinhui(Jiangxi Province Nuclear Geospatial Data Science and Systems Engineering Research Center,East China University of Technology,Nanchang 330013,China;School of Information Engineering,East China University of Technology,Nanchang 330013,China;Ganzhou Good Friends Technology Co.,Ltd.,Ganzhou 341000,Jiangxi,China)

机构地区：[1]东华理工大学江西省核地学数据科学与系统工程技术研究中心,南昌330013 [2]东华理工大学信息工程学院,南昌330013 [3]赣州好朋友科技有限公司,江西赣州341000

出　　处：《有色金属(选矿部分)》2025年第3期52-58,共7页Nonferrous Metals(Mineral Processing Section)

基　　金：江西省主要学科学术和技术带头人培养计划项目(20225BCJ22004)。

摘　　要：目前矿石分选领域使用的深度学习模型均为结构复杂、层数较深的大模型,然而大模型具有训练数据量大,对硬件设备要求高等缺点。轻量化模型不仅训练数据量远低于大模型,而且对硬件设备要求较低,针对大模型的上述问题是一种较好的解决方案。本文便致力于建立一种基于双能X射线透射图像数据的铜矿分选轻量化改进卷积网络模型CBAL2-Net。其中训练数据为2560张铜矿石图像,在深度学习模型中属于较小的样本数量。未改进模型为自建模型,基础结构简单,其中包含两层卷积层、两层全连接层及若干激活函数和池化层。自建模型的准确率在84.94%左右,且存在明显的过拟合现象。改进模型中引入CBAM轻量化注意力模块,在几乎不增加模型复杂度的情况下准确率提高了3.44%,但过拟合现象仍然存在。在本模型中使用L2正则化技术可以有效缓解过拟合的同时进一步提升识别精度。经过一系列试验验证,改进模型CBAL2-Net的分选准确率达到了93.69%。在较低的模型复杂度和资源消耗的前提下拥有较高的准确率。The deep learning models currently used in the field of ore sorting are large models with complex structures and deep layers.However,large models have the disadvantages of large amounts of training data and high requirements for hardware equipment.The lightweight model not only has much lower training data volume than the large model,but also has lower hardware requirements.It is a better solution to the above problems of the large model.This article is dedicated to establishing a lightweight improved convolutional network model CBAL2-Net for copper ore sorting based on dual-energy X-ray transmission image data.The training data consists of 2560 copper ore images,which is a relatively small number of samples in the deep learning model.The unimproved model is a self-built model with a simple basic structure,which includes two layers of convolutional layers,two layers of fully connected layers,and several activation functions and pooling layers.The accuracy of the self-built model is around 84.94%,and there is obvious over-fitting phenomenon.The CBAM lightweight attention module was introduced into the improved model,which increased the accuracy by 3.44%without increasing the complexity of the model,but the overfitting phenomenon still exists.The use of L2 regularization technology in this model can effectively alleviate overfitting and further improve the recognition accuracy.After a series of experimental verifications,the classification accuracy of the improved model CBAL2-Net reached 93.69%.It has higher accuracy with lower model complexity and resource consumption.

关 键 词：深度学习 CBAM 注意力机制 双能X射线 铜矿分选 

分 类 号：TD952[矿业工程—选矿]