Research On Yarn-dyed Fabric Defect Detection And Classification Based On Convolutional Neural Network

With the improvement of the production efficiency of textile enterprises,the quality of textile products has become increasingly stringent.The traditional manual defect detection method not only consumes a lot of manpower and material resources,but also has low detection efficiency and high missed detection rate.Yarn-dyed fabric with different types of pattern and defect has brought great difficulty to fabric defect detection and classification.The detection and classification algorithms of fabric defects based on convolutional neural network were put forward in this paper,including fabric defect classification based on the modified AlexNet,fabric defect detection based on Faster R-CNN and SSD,respectively.The convolutional neural network is used to extract and fuse the defect features automatically to take place of the traditional artificial feature extractor.A large number of fabric images with defect are utilized while model training and testing.And the experimental results analysis is given out.The research mainly contains the following aspects:(1)Researching on the classification of yarn-dyed fabric defect,the defect classification model based on the modified AlexNet is proposed.The data normalization method of the AlexNet is improved by replacing the local response normalization layer with batch normalization layer.The defect images of fabric are trained and classified by means of defect features extracted from the convolution neural network.By validating on the defect image test set,the effectiveness of the modified AlexNet classification algorithm is verified.(2)The object detection algorithm of Faster R-CNN based on convolutional neural network is used to solve the problem of fabric defect classification and detection.The architecture of Fast R-CNN is composed of region proposal network RPN and Fast R-CNN.RPN is a fully convolutional network that is used to generate bounding boxes.Then Fast R-CNN applies the proposals extracted from the RPN to classify and detect the defect in the proposal.Finally,the predicted proposals are filtered by means of non-maximal suppression algorithm and sorted by the Softmax classifier.According to the experiments,the mean average precision on the test data set for the color fabric defect can reach about90%.(3)In order to achieve real-time defect detection of yarn-dyed fabric,the SSD detection algorithm based on convolutional neural network is researched.The SSD iscomposed of the basic network and additional convolution layers.Firstly,a series of fixed-size boxes are generated,and the possibility of each box corresponding to each category is calculated.Then,the final predictions are obtained by the non-maximal suppression algorithm.The use of multi-scale feature detection in SSD can not only improve the accuracy of detection,but also shorten the detection time.Experiments proves that SSD object detection network has good detection results on defective fabric,the mean average precision of detection can reach to 83%,and the detection speed is 46 FPs to ensure the real-time detection.