Research On Multi-class Classification Of Zipper Defects Based On Deep Learning

As an important part of daily life,Zippers bring grea t conveniences to human beings.In the process of zipper generation,due to equipment and labor reasons,there are often defects such as chain teeth and upper and lower ends missing.At the same time,due to the lack of enterprises,unclear division of labor,shortage of products and labor resources and so on,detaining the development of zipper industry,which makes zipper industry free from the middle and low-end market.The traditional zipper product detection and classification methods are mainly calipers and manual visual inspection.The method has low efficiency,high false rate,high cost and cannot guarantee the reliability of zipper detection and classification.The solution couldnot meet the requirements of industrialized actual production any more.Traditional methods of machine learning have some difficulties in classifying zipper defects in complex backgrounds.In the process of production,the pipeline condition,illumination problem,light source type,camera type,zipper overlap,deformation,smudge and other problems cause the zipper defect image have low resolution,blurred background,out of focus,excessive disturbance,and detection of the zipper.It have a big impact on detection and classification of zipper detects.As a branch of machine learning,deep learning can classify large-scale images,and its performance can rival humans or even better than humans,making it have a huge application space in classification.Now there is no relevant research of the deep learning method to detect and classify zipper defects.In this paper,deep learning is applied to the multi-task classification of zipper defects to improve the efficiency and accuracy of defect classification,reduce labor intensity and labor costs.The innovations are as follow:1)Using the AlexNet model ? improving it and getting an improved version of the AlexNet+ model.the AlexNet+ model adds BatchNorm operation to reduce the relative difference between datas,while smoothing the zipper defects image,speeding up the training and improving the network generalization ability.The experiment proves that AlexNet+ has lower loss rate than the classic AlexNet in the zipper defects multi-tasking classification,and the average classification accuracy is 3.5% higher.2)Using the VGGNet model,improving it and getting an improved version of the VGGNet+ model.The VGGNet+ model discards a layer of fully connected layers and adds discarding and batch normalization operations to balance the input distribution at each layer and has stronger robustness.The classification effect and the recognition stability of VGGNet+ model is better than VGGNet and the classification accuracy is improved by 4.4%.3)Using the ResNet model,improving it and getting an improved version of the ResNet+ model.The ResNet+ model uses a small convolution kernel to extract more relevant information about zipper defects,reducing interference between zipper defects.The global average pooling operation reduces the number of parameters of the fully connected layer and can better extract the defects of the zipper.The improved ResNet+ network model has the highest classification accuracy among these classification models and the average classification accuracy is 4.5% higher than the classic ResNet network.For the conventional value,the average classification accuracy of the ResNet+ network reached 92.7%,and the value of the reached 94.6% when the value was convex.In summary,the proposed three improved models have higher classification accuracy than these original ones.The ResNet+ model can be used for multi-task classification of zipper defects types,pull tab positions,slider position and zipper position angles in actual production zipper defects data sets,and explores the multi-task classification of deep learning for zipper defects.