基于改进ResNet50的钨矿石双能X射线图像分选方法  

Dual energy X⁃ray image sorting method based on improved ResNet50 for tungsten ore

作　　者：刘志锋[1] 曾灵锋 彭芳伟 魏振华[1] 张寰宇 LIU Zhifeng;ZENG Lingfeng;PENG Fangwei;WEI Zhenhua;ZHANG Huanyu(School of Information Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013

出　　处：《现代电子技术》2024年第13期87-92,共6页Modern Electronics Technique

基　　金：江西省重大科技研发专项-揭榜挂帅项目:基于单光子效应的X射线智能选矿探测器及其应用研究(20224AAC01012)。

摘　　要：文中提出一种基于深度扩张可分离卷积和注意力机制的残差网络模型(DWAtt-ResNet),通过实验对比表明,该模型在钨矿石双能X射线图像数据集上准确率、F1分数、AUC值和AP值均优于ConvNeXt、DenseNet121和EfficientNet_b4等主流的图像分类模型。通过消融实验表明,该模型准确率达到87.4%,计算量为2.7GFLOPs,参数量为16.95M,相比ResNet50准确率提高3%,计算量降低1.42 GFLOPs,参数量降低6.56M,准确率提升的同时,效率大幅提升,更适合工业生产的矿石快速分拣需求。A residual network model based on depthwise separable dilated convolutions and attention mechanism(DWAtt-ResNet)is proposed.Comparative experiments show that the proposed model outperforms mainstream image classification models of ConvNeXt,DenseNet121 and EfficientNet_b4 in terms of accuracy rate,F1-score,AUC(area under ROC curve)value and AP(average precision)value on the dual energy X-ray image dataset of tungsten ore.Ablation experiments show that the accuracy rate of the proposed model reaches 87.4%,with a computational load of 2.7 GFLOPs and a parameter quantity of 16.95M.In comparison with ResNet50,its accuracy rate increases by 3%,its computational load decreases by 1.42 GFLOPs,and its parameter quantity decreases by 6.56M.Its efficiency is improved significantly while its accuracy rate is improved,which makes it more suitable for the rapid sorting needs of industrial production ores.

关 键 词：钨矿石 双能X射线 图像分类 ResNet50 深度扩张可分离卷积 注意力机制 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP399[电子电信—信息与通信工程]