一种优化的MobileNet模型在钼矿识别中的研究  

Research on an optimized MobileNet model for molybdenum ore recognition

作　　者：郭乾明 周才英[1] 占新龙 叶晓朗 魏远旺 GUO Qianming;ZHOU Caiying;ZHAN Xinlong;YE Xiaolang;WEI Yuanwang(School of Science,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi,China;Key Laboratory of Multi-modal Perception and Intelligent System,Jiaxing University,Jiaxing 314001,Zhejiang,China)

机构地区：[1]江西理工大学理学院,江西赣州341000 [2]嘉兴大学全省多模态感知与智能系统重点实验室,浙江嘉兴314001

出　　处：《有色金属科学与工程》2025年第2期287-296,共10页Nonferrous Metals Science and Engineering

摘　　要：针对目前仍然存在部分矿山筛选尾矿是工人手选的现状,本文提出了一种基于优化MobileNetV2模型的深度学习钼矿识别方法,该方法提升了钼矿在X射线照射下所得的灰度图像的识别精度与效率。构建了一个自标注的钼矿灰度图像数据集,并且对图像进行了预处理和归一化。在MobileNetV2架构的基础上,进行了创新和改进,引入了坐标注意力机制(Coordinate Attention,CA),通过调整宽度因子和L2正则化参数,在增强模型特征提取能力和泛化能力的同时减少了训练时间。实验结果表明,与原始的MobileNetV2模型相比,本方法在钼矿识别任务上的准确率提升了3.5%,同时训练时间得到了显著减少,与几种典型卷积神经网络架构如ResNet50、EfficientNetB0、VGG16等相比,本模型在准确率、参数量、训练时间等多个关键指标上均展现出了显著优势。Given the current situation where most of the tailing screening in mines is manual selection by workers,this paper proposed a deep learning molybdenum ore recognition method based on an optimized MobileNetV2model,which would improve the recognition accuracy and efficiency of molybdenum ore in gray-scale images obtained under X-ray irradiation.A self-labeled molybdenum ore gray-scale image data set was constructed,and the images were preprocessed and normalized.Based on the MobileNetV2 architecture,innovations and improvements were made by introducing the coordinate attention mechanism(Coordinate Attention,CA).Adjusting the width factor and L2 regularization parameters enhanced the model's feature extraction ability and generalization ability while reducing the training time.The experimental results showed that compared with the original MobileNetV2model,the accuracy of this method in the molybdenum ore recognition task was increased by 3.5%.At the same time,the training time was significantly reduced.Compared with several typical convolutional neural network architectures such as ResNet50,EfficientNetB0,and VGG16,this model showed significant advantages in multiple key indicators such as accuracy,number of parameters,and training time.

关 键 词：钼矿识别 MobileNetV2 CA注意力机制 宽度因子 L2正则化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]