基于CaffeNet的工业瓶口缺陷检测  

Defect Detection of Industrial Bottle Mouth Based on CaffeNet

作　　者：张良 张卫华[2] 周激流[1] Zhang Liang;Zhang Weihua;Zhou Jiliu(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065;College of Computer Science,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学电子信息学院,成都610065 [2]四川大学计算机学院,成都610065

出　　处：《现代计算机》2022年第1期101-105,120,共6页Modern Computer

基　　金：国家重点研发计划重点专项(司法专题任务)(2018YFC0830300)。

摘　　要：在陶瓷药瓶生产过程中,容易出现瓶口破碎或不完整等情况。为解决陶瓷瓶口的缺陷检测问题,本文提出了一种基于卷积神经网络CaffeNet模型的陶瓷瓶口分类方法。该方法利用工业生产的瓶口残次品和正品建立样本数据集,利用卷积神经网络的卷积层和池化层对残次品、正品的图像特征进行提取,通过求得最小的交叉熵损失来获取适配本样本数据集的最佳权重,已达到较好的分类效果。此外,在给定相同数量测试数据集的情况下,使用CPU、GTX 1060 Mobile、GTX Titan X Pascal和华为Atlas200DK对给定数据集进行分类所需要的功耗进行对比。实验结果表明,在使用相同的权重文件的情况下,对一定数量的测试数据集进行分类,华为Atlas200DK在达到97%以上准确率的同时,所需时间与GTX Titan X Pascal持平,功耗达到最低,可为工业化检测药瓶缺口提供一种高效的、低能耗的方案。In the production process of ceramic medicine bottle,the bottle mouth is easy to be broken or incomplete.In order to solve the defect detection problem of ceramic bottle mouth,this paper proposes a classification method of ceramic bottle mouth based on convolution neural network caffenet model.This method used industrially produced defective bottles and genuine products to establish a sample data set,and used the convolutional layer and pooling layer of the convolutional neural network to extract the image features of the defective and genuine products,the optimal weight of the sample data set was obtained by minimizing the cross entropy loss,and the better classification effect has been achieved.In addition,the power consumption of CPU,GTX 1060 Mobile,GTX Titan X Pascal and Huawei Atlas 200 DK to classify the given data set was compared when the same number of test data sets are given.The experimental results show that when using the same weight file to classify a certain number of test data sets,the accuracy of Huawei atlas 200 is more than 97%,the time required is the same as GTX Titan X Pascal,and the power consumption is the lowest.It can provide an efficient and low-energy scheme for industrial detection of drug bottle gap.

关 键 词：深度学习 瓶口检测 低功耗 

分 类 号：TQ174.7[化学工程—陶瓷工业] TQ460.69[化学工程—硅酸盐工业] TP391.41[自动化与计算机技术—计算机应用技术]