Research And Application Of Deep Neural Network In Defect Detection Of Yarn-dyed Shirts

Yarn-dyed shirt with beautiful pattern is one of the garment products with the strongest ability to earn foreign exchange through export,earning more than US $6 billion annually.However,with the increasing demand of customization and small batch market,flexible manufacturing technology,yarn-dyed shirts can switch between mass production,small batch production and customization production,it has become an important problem to be solved urgently.Among them,the automatic defect detection of yarn-dyed shirt is an important part of the flexible manufacturing problem in the process of yarn-dyed shirt production.The quality of yarn-dyed shirt is seriously affected by the defects of the pieces,but there is no effective method to detect the defects of the pieces.In this paper,a new method of defect detection for yarn-dyed shirt using depth convolutional neural network is presented.A pre-experiment on pattern classification of dyed fabrics was carried out by using classic convolutional neural network AlexNet and GoogLeNet,which proved the validity of pattern feature extraction from convolutional neural network,on this basis,the defect detection algorithm model based on RetinaNet,Faster RCNN and YOLOv2 is further developed.In this paper,a large-scale image database of yarn-dyed shirts segment defects is established,and then,based on this database,yarn-dyed shirts segment defects model based on RetinaNet,Faster RCNN and YOLOv2 is constructed,which can be used to analyze the defects of yarn-dyed shirt segments,Finally,the defect detection model is tested on GPU platform and embedded platform respectively.The research contents mainly include the following aspects:(1)Because of the problem of high error detection rate of traditional image feature engineering and shallow-layer machine learning,a defect detection model based on RetinaNet is established in this paper.The model uses an improved cross-entropy Loss function named Focal Loss and incorporates the depth residual network model structure.(2)Because of the problem of the low precision of yarn defect detection,an algorithm based on the network model of Faster RCNN is proposed.Firstly,the feature of yarn-dyed shirt is extracted by combining the algorithm frame with VGG16 network model,then the candidate regions are predicted by using RPN region,and finally the candidate regions are detected and classified.The final model improves the detection accuracy of belt yarn defects.(3)In order to improve the real-time performance of RetinaNet and Faster RCNN models,an improved YOLOv2 network-based defect detection algorithm for yarn-dyed shirt is proposed.By selecting the initial model,expanding the training data,choosing the training batch and the cutting scale,optimizing the clustering method,matching the model with the anchor box,and adjusting the structure of the final network model,an optimization model for defect detection of yarn-dyed shirt pieces is obtained.The model is superior to RetinaNet and Faster RCNN in the performance of defect detection.The model can realize defect training and real-time detection on GPU platform.The optimized YOLOv2 yarn-dyed shirt cutting defect detection model is transplanted to Nvidia Jetson TX2 embedded AI hardware platform,and the real-time testing results of yarn-dyed shirt cutting defect detection are also obtained.Finally,the feasibility of the terminal deployment of the depth-learning model for the defect detection of the yarn-dyed shirt is verified.(4)Summarize the completed research work,and look forward to the future development of the subject.Figure 27,table 15,53 references.