融合交叉熵损失的3DCNN探水作业动作识别  

Action recognition for water exploration based on 3DCNN combined with cross-entropy error

作　　者：刘春霞[1] 高强 潘理虎[1] 龚大立 LIU Chun-xia;GAO Qiang;PAN Li-hu;GONG Da-li(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China;Jingying Shuzhi Technology Limited Company,Taiyuan 030006,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024 [2]精英数智科技股份有限公司,山西太原030006

出　　处：《计算机工程与设计》2022年第4期1160-1165,共6页Computer Engineering and Design

基　　金：山西省应用基础研究计划基金项目(201901D111252);先进控制与装备智能化山西省重点实验室开放课题基金项目(ACEI202002)。

摘　　要：为解决矿井探水作业中人工验收效率低、耗时长等问题,提出一种融合交叉熵损失函数的3DCNN探水作业动作识别模型。使用ReLU非线性化函数和SoftMax交叉熵损失函数,将隐含的特征数据判断分类别后再进行学习,得到较为丰富的批次网络信息图;将批量归一化操作引入到网络结构中,弥补网络模型收敛速率慢的不足,提高模型的泛化能力和鲁棒性,达到有效提高验收效率的目的。经过与其它网络模型对比,实验结果表明,该方法有效解决了人工验收低效率的问题,提高了动作识别的准确率。To solve the problems of low efficiency and time-consuming manual acceptance in mine water exploration operations,the model of action recognition for water exploration based on 3DCNN combined with cross-entropy error function was proposed.The ReLU nonlinearization function and the SoftMax cross-entropy loss function were used to classify the implied feature data,and the batch network information graph was obtained.The batch normalization operation was introduced into the network structure to make up for the slow convergence rate of the network model,so as to improve the generalization and robustness of the model.The purpose of effectively improving the efficiency of acceptance was achieved.Comparing with other network mo-dels,the experimental results show that the proposed method effectively solves the problem of low efficiency in manual accep-tance and significantly improves the accuracy of action recognition.

关 键 词：煤矿水害防治 探水作业 图像处理 行为分类 深度学习 三维卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]