Detection Classifiation Of Cigarette Appearance Defects Based On Deep Learning

China is a big tobacco country,and cigarettes are one of the important products of tobacco.According to statistics,the current number of smokers in China is roughly 350 million;on the other hand,the annual tax paid through tobacco is also trillions.In 2020,its industrial and commercial tax profits will reach more than 1.2 trillion yuan.Yunnan Tobacco is an important producer of tobacco in China,and its output ranks first in the country.For Yunnan Province,Yunnan Tobacco is undoubtedly one of the most core economic industries.In the process of cigarette manufacturing line,various reasons,such as machine failure and damage to raw materials,will cause quality problems in cigarettes,resulting in different types of appearance defects.In the past,only manual detection and some traditional digital image processing methods were used for detection.These detection methods will inevitably bring some problems:(1)manual detection will inevitably bring errors and slow speed;(2)digital image processing methods It is impossible to distinguish all types of defects,and the identified defects cannot be specifically classified,which leads to the inability to timely feedback to the production end,thereby failing to reduce production costs.Aiming at the above two problems,this paper does the following research:Design a method for detecting appearance defects of cigarettes based on the improved YOLOv4(You Only Look Once Version 4)model.This method first enhances the original data set and improves the generation method of the prior frame,and then introduces the attention mechanism on the YOLOv4 model,and replaces it with the Atrous Spatial Pyramid Pooling structure of ASPP(Atrous Spatial Pyramid Pooling)The original Spatial Pyramid Pooling(Spatial Pyramid Pooling)structure further improves the accuracy of cigarette appearance defect detection.The mAP of the improved algorithm reaches 91.77%,which is 4.88%higher than the original method,the Precision reaches 93.32%,which is 5.45%higher than the original method,and the Recall reaches 88.81%,which is 4.05%higher than the original method,and the FPS drops by 7 frames.But this does not affect the actual production line process.The experimental results show the effectiveness of the improved method in the detection of cigarette appearance defects.In actual industrial production,it is inevitable to encounter image classification problems.Therefore,based on the image data of the appearance of cigarettes,this paper applies the deep learning method to the image classification problems encountered in the actual industrial production process.This paper studies the image classification of appearance defects of cigarettes,and strives to make a little contribution to the subsequent classification requirements of cigarette factories and other image classification problems encountered in the industrial production process.(Split-Attention Networks)model for the classification of appearance defects of cigarettes.First,to solve the problem of insufficient samples of cigarette appearance defects,the transfer learning method is adopted;secondly,according to the characteristics of cigarette images,multi-scale training is adopted,and pictures of different sizes are input for training;finally,in order to better extract Defect features,improve the classification accuracy,and replace the ReLU activation function with h-swish.The accuracy and recall rate of the improved network are 92.16%and 90.79%,respectively,which are 6.12%and 5.32%higher than the original method,and both the accuracy and recall rate are higher than the current mainstream classification networks.