Research On Cigarette Box Defect Detection Based On Deep Learning

With the development of modern industry,the demand of realizing production automation is higher and higher.In the production and manufacturing process of cigarette case,every good cigarette case needs to be detected manually,and defective products such as glue spots,dents and degummings are removed.The cigarette pack production line can produce tens of thousands of cigarettes a day,so the manual inspection is time-consuming and laborious.If computer vision can replace manual inspection,it will make the manufacturing process more intelligent,reduce the artificial resources,and achieve the effect of half power.Over the past few years,deep learning technology has advanced dramatically,and has been applied in many fields.It has replaced many tasks that need to be done manually,and its effect is equal to or even better than that of manual work.Therefore,this study will study how to detect cigarette package defects from the direction of deep learning target detection,instead of traditional manual detection,so as to reduce costs,improve product quality,detect accuracy and reduce the rate of defective products.At the same time,it also makes a certain contribution to the research of other packaging defects detection.The main research content of this paper includes the following five parts:(1)The research of domestic and foreign scholars on defect detection was reviewed.Due to the difficulty of defect detection and the requirements of subsequent hardware deployment,data enhancement and model lightweight were also studied,laying a theoretical foundation for the work of this paper.(2)Field investigation of the production line of Chinese cigarette packs,photographing and obtaining the data set of Chinese cigarette packs with defects,and labeling them,obtaining labeling documents,making preparations for the subsequent study of cigarette pack defect detection.Different deep learning target detection algorithms such as Center Net,YOLOv3 and YOLOX algorithms are studied,and these algorithms are used to train cigarette box data sets.By analyzing and comparing the detection accuracy,speed and defect detection effect of the algorithm,YOLOv3 algorithm is used as the basis for optimization and improvement.Through the sample distribution of the data set and the training results on YOLOv3,it is found that the data set has some problems such as uneven distribution of defect samples and low accuracy of glue spot detection.In view of the above two kinds of problems,from the two perspectives of offline data enhancement and online data enhancement,the methods of over-sampling and spoon-feeding data enhancement are used to solve the problems effectively.(3)Studied how to improve the m AP of average detection accuracy of Carton-YOLO defect detection algorithm.By analyzing the size of data set images and the influence of input images on algorithm detection,the size of network input is optimized to increase the size of input images and enlarge the defects of cigarette boxes,making it easier for the model to extract features and thus improve m AP.Then,the loss function is optimized from two aspects: on the one hand,the loss weight of small defects is improved;on the other hand,the model pays more attention to positive samples rather than the background of images.In addition,the original IoU is improved to CIoU,which greatly improves the convergence speed,saves the cost of model training and improves m AP.(4)How to compress the cigarette pack defect detection model so that it can be deployed on embedded devices is studied.By studying the differences in the ways and computational quantities of deeply separable convolution and ordinary convolution,the convolution modes were changed by region and experiments were carried out.After observing the results,part of ordinary convolution was replaced and part of model size was reduced.At the same time,the model quantization method is adopted.After the training,the model is quantized statically,so that the weight represented by float32 is represented by INT8 through the quantization tool in Py Torch,which makes the model size 3.78 times smaller than before.(5)The interface of cigarette case surface defect monitoring system is introduced.By comparing and analyzing the advantages and disadvantages of interface development tools Py Qt,Tkinter and WXPython,combined with the development requirements of this system,WXPython framework is selected to design and develop the interface of the monitoring system.Then,the login interface,monitoring interface,data interface and functional modules in the interface are designed and introduced.This research work dedicated to the fulfillment of the cigarette case of intelligent defect detection task,replace the artificial vision with computer vision,supplemented by the target detection algorithm and intelligent hardware,deploy them in cigarette production line,make it easier for cigarette case in the process of production defect detection,and no need to remove from the production line of cigarette packets for defect detection.This study provides some guidance for the real-time monitoring of cigarette box production line and related systems and key technologies of cigarette box defect detection in cigarette box defect detection task,and also plays a certain reference role for other packaging box defect detection research.