基于注意力增强YOLOv5l的矿粉品位识别算法优化研究  

Optimization Study of Ore Grade Recognition Algorithm Based on Attention-enhanced YOLOv5l

作　　者：丁鹏益 徐振洋 郭连军 王雪松 DING Pengyi;XU Zhengyang;GUO Lianjun;WANG Xusong(School of Mining Engineering,University of Science and Technology Liaoning,Anshan 114051,China;Engineering Research Center of Green Mining of Metal Mineral Resources Liaoning Province,Anshan 114051,China;School of Architecture and Civil Engineering,Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]辽宁科技大学矿业工程学院,辽宁鞍山114051 [2]辽宁省金属矿产资源绿色开采工程研究中心,辽宁鞍山114051 [3]沈阳工业大学建筑与土木工程学院,辽宁沈阳110870

出　　处：《金属矿山》2024年第11期151-157,共7页Metal Mine

基　　金：国家自然科学基金项目(编号:51974187)。

摘　　要：现阶段传统化学分析方法获取品位存在费时费力等问题,通过图像识别分析块状矿石品位又存在形状干扰严重的问题,为此提出了一种基于YOLOv5针对矿石矿粉特征的图像识别方法。同时,在训练过程中添加卷积注意力模块(Convolutional Block Attention Module,CBAM)、挤压和激发模块(Squeeze and Excitation Module,SENet)进入训练网络,通过注意力机制增强学习矿粉具体特征的能力,将注意力聚焦于矿粉明显细节,忽略矿粉图中无用信息,提高识别精度;其次通过修改损失函数增强其分类效果,研究损失函数对矿粉识别效果的影响。研究表明:在铁矿粉品位识别中,添加CBAM注意力模块的网络模型识别矿粉的训练精度达到86%,使用SENet注意力模块的网络训练精度为80%,均略高于原有模型的79%,修改损失函数的网络模型训练精度降低了5%,得出YOLOv5l+CBAM且损失函数设置为0.5的网络模型最佳。研究结果反映出所提方法对矿粉特征图像识别具有一定的适用性。At the current stage,traditional chemical analysis methods for grade determination have been found to be timeconsuming and laborious.Similarly,image recognition for grade analysis of bulk ore suffers from significant shape interference.To address these issues,an image recognition method for ore and ore powder features based on YOLOv5 is proposed.During the training process,the Convolutional Block Attention Module(CBAM)and the Squeeze and Excitation module(SENet)into the neural network to enhance the model′s ability to learn specific features of ore powder has been incorporated.Significant details in the ore powder were focused towards by the attention mechanism while ignoring irrelevant information,thereby improving recognition accuracy.Additionally,the loss function to enhance its classification effect,and investigate the impact of the loss function on the effectiveness of ore powder recognition were modified.The study results show that in the grade recognition of iron ore powder,the network model with the added CBAM attention module has a training accuracy of 86%,and the network model using the SENet attention module has a training accuracy of 80%.Both are slightly higher than the original model′s accuracy of 79%.However,the network model with the adjusted loss function saw a decrease in training accuracy by 5%.The YOLOv5l+CBAM model with a loss function set to 0.5 is optimal was concluded.The study results demonstrate that the proposed image recognition method for ore powder features has certain applicability.

关 键 词：铁矿 矿石品位 YOLOv5 图像识别 卷积注意力模块 挤压和激发模块 

分 类 号：TD672[矿业工程—矿山机电]