基于改进MobileNet V3的矿物智能识别模型  

Intelligent Recognition Model of Mineral Based on Improved MobileNet V3

作　　者：宛鹤[1] 张金艳 屈娟萍 张崇辉[1] 薛季玮 王森[1] 卜显忠[1] WAN He;ZHANG Jinyan;QU Juanping;ZHANG Chonghui;XUE Jiwei;WANG Sen;BU Xianzhong(School of Resource Engineering,Xi′an University of Architecture and Technology,Xi′an 710055,China;Oulu Mining School,University of Oulu,Oulu FI-90014,Finland)

机构地区：[1]西安建筑科技大学资源工程学院,陕西西安710055 [2]奥卢大学奥卢矿业学院,芬兰奥卢FI-9004

出　　处：《金属矿山》2024年第1期174-181,共8页Metal Mine

基　　金：国家自然科学基金项目(编号:52274271,52074206,52104266)。

摘　　要：针对当前矿物识别领域存在的精度不佳、适应性差、携带不便等问题,提出了一种基于改进MobileNet V3的矿物智能识别模型(CA-MobileNet V3)。为获得研究所需的有效数据集,通过由mindat. org网站和自行拍摄方式获取的矿物图像创建了一个包含19种矿物的数据集,对其进行数据增强处理,并按照8:1:1的比例划分为训练集、验证集和测试集。为提升模型对图像信息的特征提取能力,引入协调注意力机制,用以替代轻量型MobileNet V3模型的原始SE注意力机制,以提高矿物识别准确率。最后,采用迁移学习方法预训练CA-MobileNet V3模型,以加速模型收敛、提高泛化能力、避免过拟合。在训练过程中,将CA-MobileNet V3与mobilenet v3、MobileNet V3、ShuffleNet V2、Efficient Net V2等模型进行了性能比较。结果表明:各迁移模型均展现出显著的收敛速度优势,而CA-MobileNet V3矿物智能识别模型的Top1-准确率、Top2-准确率、f_1-score值分别达到93.90%、98.58%和93.89%,在所有模型中效果最佳,且模型大小仅为4.61 MB,属于轻量化模型。为验证模型有效性,t-SNE可视化分析被用于不同模型的识别效果比较,进一步印证了CA-MobileNet V3模型的优越性。Challenges such as low accuracy,limited adaptability,and a lack of portability in the field of mineral recogni-tion were addressed through the proposal of an intelligent mineral recognition model(CA-MobileNet V3)based on the im-proved MobileNet.For research purposes,a collection of 19 minerals was compiled from images sourced from mindat.org and self-captured images.The images,following data enhancement procedures,were categorized into training,validation,and test sets in an 8∶1∶1 ratio.To enhance the feature extraction capabilities,the original SE attention mechanism in the lightweight MobileNet V3 model was replaced by the coordination attention mechanism.The aim of this alteration was to elevate mineral recognition accuracy.Subsequently,pre-training using transfer learning was employed on the CA-MobileNet V3 model to expe-dite convergence,enhance generalization,and mitigate overfitting.During training,a comparison of the performance of CA-Mo-bileNet V3 was made with other models,including mobilenet v3,MobileNet V3,ShuffleNet V2,and EfficientNet V2.The results revealed notable advantages in terms of convergence speed for all transfer models.Particularly,the CA-MobileNet V3 model a-chieved TOP1-accuracy,TOP2-accuracy,and f1-score values of 93.90%,98.58%,and 93.89%,respectively,showcasing su-perior performance compared to other models.Furthermore,this lightweight model boasted a compact size of only 4.61 MB.To further validate the model′s effectiveness,t-SNE visual analysis was employed,providing a comparative assessment of the recog-nition effects among different models.

关 键 词：矿物分类 迁移学习 轻量化模型 协调注意力机制 t-SNE 

分 类 号：P578[天文地球—矿物学] TD67[天文地球—地质学]