基于卷积神经网络特征提取的轻量级包装袋分类模型  

Lightweight Bag Classification Model Based on Convolution Neural Network Feature Extraction

作　　者：鲁江坤 汪林林[2] 陈红阳[1] LU Jiang-kun;WANG Lin-lin;CHEN Hong-yang(School of Computer Engineering,Chongqing College of Humanities,Science and Technology,Chongqing 401524,China;School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆人文科技学院计算机工程学院,重庆401524 [2]重庆邮电大学软件工程学院,重庆400065

出　　处：《塑料科技》2020年第8期69-72,共4页Plastics Science and Technology

基　　金：重庆市教委科技项目(KJ1716367);重庆市教委科技项目青年项目(KJQN201801801)。

摘　　要：为解决塑料包装袋难识别以及难收集问题,应用神经网络提出一个轻量级包装袋分类模型。模型网络分为预处理层、特征提取层和分类层,且在预处理层模拟手工特征提取方式,应用SQUARE 3×3高通滤波器捕获丰富图像纹理特征。经仿真研究模型各层卷积核数量、卷积核尺寸、训练时长、批次样本数对模型的影响,方案6各层卷积核数量、尺寸和批次样本数分别为64、3×3和8,捕获丰富包装袋特征,具备最佳测试分类准确率0.7305,较适用于实际包装袋分类工作,高于传统HOG模型约13%。实验结果可为当前包装袋识别分类工作提供一定的参考。In order to solve the problem of difficult identification and collection of plastic bags,a lightweight bag classification model was proposed by using neural network.The model network was divided into preprocessing layer,feature extraction layer and classification layer.In the preprocessing layer,the manual feature extraction method was simulated,and the square 3×3 high pass filter was used to capture image texture features.The number of convolution kernels,the size of convolution kernels,the time of training and the number of samples in each layer of the model were studied by simulation.In scheme 6,the number and size of convolution kernels,and batch samples number in each layer were 64 and 3×3 respectively,which could capture the characteristics of a variety of bags.It got the best test classification accuracy of 0.7305,which was more suitable for the actual bag classification work,about 13%higher than the traditional HOG model.The experimental results can provide some reference for the current work of bag recognition and classification.

关 键 词：塑料 手工提取 轻量级 深度学习 

分 类 号：TU532+61[建筑科学—建筑技术科学]