Printing Product Defect Detection Based On Deep Learnin

With the rapid development of the printing industry,more strict standards have been put forward for the quality of printed matter.However,at present,printed matter defect detection in our country is in the medium stage of development.In view of the shortcomings of low detection rate and poor real-time performance of printed matter in embedded devices,this paper has done the following three aspects of work based on deep learning.The main algorithms in this paper include:(1)Aiming at the characteristics of printed matter defects,this paper proposes a detection model based on YOLOv4-Tiny.Firstly,in order to effectively reduce noisy data,an improved attention module is introduced into the backbone network of YOLOv4-Tiny to enhance the model’s ability to focus on the region.Secondly,in order to improve the effect of multi-scale feature fusion,bicubic interpolation is implemented in the up-sampling module.In order to solve the redundancy detection results of the region boundary frame,an optimized non-maximum suppression method is proposed to make the detection results better correspond to the region border.(2)To reduce the detection time of YOLOv4-Tiny.Based on YOLOv4-Tiny model,two ResBlock-D modules in ResNet-D network are used,and CSPBlock module is deleted,thus reducing algorithm complexity and improving detection speed.(3)An auxiliary network module is designed to extract more feature information of objects and increase attention to details.In the design of the auxiliary network,we use two consecutive 3×3 convolutional layers and add channel attention and space attention at the same time,so that the model can pay attention to more information that should be paid attention to and reduce the interference of useless information.Finally,the auxiliary network and backbone network are integrated to construct the whole network structure of the improved YOLOv4-Tiny.Compared with the original algorithm,the average accuracy and FPS of the improved algorithm are increased by 6.8% and 18,reaching 84.2% and 289.Finally,a defect detection system is designed,which realizes the functions of system login,image acquisition,defect detection and so on.The experiment shows that the target is achieved,the precision is met,and it has a certain practical value for print defect detection.