Research On The Detection Method Of The Defect Of The Mug Mouth Based On Deep Learning

Mugs are favored by consumers for their diverse appearance.In the manufacturing process,under the influence of objective factors such as manufacturing process and manufacturing environment,defects such as gap,scratch and speckle may appear at the mouth of the mug.At present,in the production line of mug,artificial quality inspection is mainly used to detect.Such detection method is limited by human identification ability and has certain limitations.For this reason,this thesis adopts the detection method of the defect of the mug mouth based on deep learning,and the main research contents are as follows:Firstly,build the data set of the defects of the mug mouth and Analysis the defect detection algorithm,Faster R-CNN is selected as the defect detection algorithm.The working principle and training strategy of Faster R-CNN are introduced.Secondly,according to the loss of training,resnet-50 is selected as the feature extraction network,and the Faster R-CNN algorithm is used to detect the defects of the mug mouth.The analysis results show that the algorithm is not effective in detecting small defects.Finally,aiming at the problem that the detection effect of small defects is not good,we combine Faster R-CNN with feature pyramid networks(FPN)to make better use of the features of low-level networks,and integrate the high-level features with rich semantic information and the low-level features with highresolution,so as to improve the detection precision of small defects.Aiming at the imbalance of positive and negative sample distribution in RPN training,focal loss is used to improve the cross entropy loss function.It not only reduces the impact of the imbalance of the number of samples,but also reduces the impact ofa large number of easily divided negative samples on the back propagation gradient in the training process.In this thesis,after the study of the above methods,we trained the samples of the defects in the mouth of the mug on Faster R-CNN,Faster R-CNN+ FPN,Faster R-CNN+Focal loss and Faster R-CNN+FPN+Focal loss respectively.After training,the test set is sent to the model for detection,and the detection precision of the improved model is better than that of the original Faster R-CNN,especially the Faster R-CNN+FPN+Focal loss,with the detection mean Average Precision reaching 85.4%,5.2% higher than that of the original Faster R-CNN.Among them,for the speckle defect,the detection Average Precision is improved by 9.1%.