摘要
基于深度Inception网络的包装袋检测模型,设计3个Inception模块,称为Inception1、Inception2和Inception3;设计多个方案开展检测性能研究工作。经仿真分析,方案七包装袋检测模型拥有8个网络层,在第1~3个、第7个网络层应用卷积核尺寸为3×3的标准卷积,在第4~6个网络层依次应用Inception1、Inception2和Inception3模块,获得最佳的检测准确率0.7684。说明较前网络层的Inception1模块可捕获细致的塑料包装袋纹理特征;Inception2模块应用多尺寸卷积核和平均池化可捕获多样化纹理特征;较末端网络层的Inception3模块可对高度抽象特征进行特征变换和提取。方案七模型检测准确率高于传统HOG模型约12%。
Based on the deep Inception network packaging bag inspection model,design three Inception modules,called Inception1,Inception2 and Inception3;design multiple programs to carry out inspection performance research work.Through simulation analysis,the seven-pack inspection model has 8 network layers,and standard convolutions with a convolution kernel size of 3×3 are applied in the 1 st to 3 rd and 7 th network layers,and in the 4 th to 6 th network layers.Apply the Inception1,Inception2 and Inception3 modules in turn to obtain the best detection accuracy of 0.7684.Explain that the Inception1 module of the previous network layer can capture detailed texture features of plastic bags;the Inception2 module can capture multiple texture features by applying multi-size convolution kernels and average pooling;the Inception3 module of the end network layer can perform highly abstract features Feature transformation and extraction.The detection accuracy rate of the model in Scheme 7 is about 12%higher than that of the traditional HOG model.
作者
韦超英
李海强
凌志梅
WEI Chao-ying;LI Hai-qiang;LING Zhi-mei(Guangxi City Vocational College,Chongzuo 532200,China)
出处
《塑料科技》
CAS
北大核心
2020年第7期60-63,共4页
Plastics Science and Technology
