基于改进ResNet50在高岭土矿石图像分类的应用  

Application of Improved ResNet50 in Image Classification of Kaolin Ores

作　　者：刘凡 陈锐 Liu Fan;Chen Rui(School of Information Engineering,East China University of Technology,Nanchang 330013,China;Jiangxi Province Engineering Research Center of New Energy Technology and Equipment,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学机械与电子工程学院,南昌330013 [2]东华理工大学江西省新能源工艺及装备工程技术研究中心,南昌330013

出　　处：《机电工程技术》2024年第5期167-171,共5页Mechanical & Electrical Engineering Technology

基　　金：国家自然科学基金(12165001);江西省科技厅重大科技研发专项(20224AAC01012)。

摘　　要：煤矸石是生产高岭土的直接原料。为了能快速精准地将含有高品位高岭土的煤矸石从原始矿石中预分选出来,提出了一种改进ResNet50的小模型图像分类识别方法。基于X射线与物质的相互作用原理,对已划分的高低品位原生矿石进行透射成像(灰度图),使用labelImg标签软件制作数据标签;基于Pytorch深度学习框架,使用数据增强算法扩增样本容量;基于ResNet50残差网络架构,使用多尺度可分离卷积降低卷积计算量并加深网络深度,采用双通道池化替代单一池化来均衡图像局部特征,优化残差结构和微调超参数使网络学习性能达到更优。结果表明:在相同的实验条件下,与传统的ResNet18/34/50和VGG16网络相比,改进的ResNet50网络在验证集上分类准确率最高,达到97.87%;对比GoogLeNet网络,两者分类精度相近的同时,改进的ResNet50网络的整体学习速度提升了接近4倍。Gangue is the direct raw material for the production of kaolin.In order to quickly and accurately pre-sort gangue containing high-grade kaolin from the original ore,a small model image classification and recognition method based on improved ResNet50 is proposed.Based on the principle of interaction between X-rays and materials,transmission imaging(grayscale map)is performed on the classified high and low grade primary ores,and data labels are made by using labelImg labeling software.Based on the PyTorch deep learning framework,the sample size is amplified by data augmentation algorithm.Based on the ResNet50 residual network architecture,multi-scale separable convolution is used to reduce the amount of convolution calculation and deepen the network depth,dual-channel pooling is used instead of single pooling to balance the local features of the image,and the residual structure and hyperparameters are optimized to achieve better network learning performance.The results show that under the same experimental conditions,compared with the traditional ResNet18/34/50 and VGG16 networks,the improved ResNet50 network has the highest classification accuracy on the verification set,reaching 97.87%.Compared with the GoogLeNet network,while the classification accuracy of the two is similar,the overall learning speed of the improved ResNet50 network is increased by nearly 4 times.

关 键 词：煤矸石与高岭土 深度学习 图像识别 改进ResNet50 图像增强 灰度直方图 

分 类 号：TD82[矿业工程—煤矿开采] TD92[矿业工程—矿山开采] TP391.41[自动化与计算机技术—计算机应用技术]