Research On Real-time Visual Detection Of Recycled Cigarette Box

In order to save paper resources,cigarette factories need to sort and reuse a large number of recycled cigarette boxes every year.After investigation,at present,cigarette factories mostly use manual visual methods to classify and filter according to the graphic identification and damage degree of recycled cigarette boxes.There is an urgent need for a visual inspection technology that can automatically identify the type and damage of the cigarette box to replace the current manual labor.However,the graphic information on the recovered cigarette box is blurred,the contrast is low,and the scale characteristics of the damaged area are greatly different,which restricts the performance of the visual detection algorithm of the recycled cigarette box.Aiming at the limitation of manual sorting,this paper proposes to use deep learning-based target detection algorithm to detect the type and damage of recovered cigarette boxes.The main research contents are as follows:1.Use the photos of the collected cigarette boxes collected in the cigarette box recycling warehouse of Qujing Cigarette Factory to make a data set.There are five types of identification space,including four types of cigarette boxes and defect.2.Using Faster-Rcnn as the algorithm framework,applying VGG16 and Res Net50 network models to realize the classification and defect identification of recovered cigarette boxes,of which Res Net50 has the best effect,the detection accuracy of cigarette box types is 100.0%,and the defect class has reached 79.40%,m AP reached 91.40%.3.Aiming at the problems of slow detection speed of Faster-Rcnn and low accuracy of YOLOv2 algorithm defect detection,based on the network structure of YOLOv2,the characteristics of the bottom three scales are connected in parallel,build the YOLOv2-PF model,which improves damage to the original network structure Class detection accuracy.The detection accuracy of the type of cigarette box is 100.0%,the accuracy of the defect class reaches 84.96%,and the m AP can reach 96.64%.4.Aiming at the problem that the single-scale network has a poor detection effect on small damaged targets,YOLOv3 based on the feature pyramid is used to detect the recovered cigarette box,but its feature extraction part is redundant in structure,and the YOLOv3 network is simplified through the ablation experiment,build the YOLOv3-YX model,and added a coefficient to control the coordinate loss and frame loss in the loss function.The detection accuracy of the type of cigarette case is 100.0%,the accuracy of the defect class is increased to 87.96%,and the m AP is increased to 97.59%.5.Use the YOLOv3-YX model to build a human-computer interaction interface,including three functions: single frame detection,offline video detection,and online video detection.It can detect the type and damage of the target cigarette box in real time,and the video detection speed can reach 12 FPS.In this paper,four sets of cigarette boxes and their surface defect data sets are established,realize the detection and damage detection of recycled cigarette boxes,and a human-computer interaction interface is built to promote the application of the target detection algorithm based on deep learning to the real-time detection of recycled cigarette boxes Provides an effective method basis.