| With the rapid development of the textile industry,the production of cloth has become larger and larger,which has brought greater challenges to the detection of textile defects.Traditional manual inspection techniques have become increasingly difficult to meet industrial production needs.The development of computing technology,especially artificial intelligence technology,provides new and more reliable technical support for fabric defect detection.In this thesis,based on the variety of cloth defects,various scales and serious background interference,a method based on deep learning to identify and detect cloth defects is proposed.Aiming at the difficulties in the identification of cloth defects,such as various types,various scales and serious background interference,this thesis proposes a method based on deep learning to identify and detect cloth defects.Aiming at the problem of cloth defect identification,this thesis proposes a cloth classification algorithm based on improved Inception-ResNet-v2 network.According to the actual application requirements,this thesis divides the defect identification of cloth into two tasks: the two-category identification of normal cloth and defective cloth,and the multiclassification identification of normal cloth and 7 kinds of defect cloth.In order to extract the richer features in the cloth,the number of layers of the network is further increased,and the size of the convolution kernel in the network is improved.The finally improved algorithm obtained more than 94% accuracy in the cloth two classification identification,and in the multi-class identification,the top-1 accuracy rate was over 86%.At the same time,by comparing the effects of the improved algorithm before and after,the improved algorithm can improve the accuracy of cloth defect recognition by nearly 20%,and successfully verify the effectiveness of the improved algorithm.Aiming at the problem of fabric defect detection,this thesis proposes a fabric defect detection algorithm based on improved Faster-RCNN network.The main improvement points are: designing multi-scale feature map extraction network;for different kinds of cloth defects,based on k-means algorithm to design anchor points of different sizes and proportions;for the cloth defect target,there is no target inclusion,right The maximal suppression algorithm is correspondingly improved.Considering the serious inter-class similarity of cloth defects,this thesis designs a Faster-RCNN cloth defect detection algorithm based on pre-multi-classifier,and finally achieves defect detection mAP@IoU=0.5 up to 82%.The effectiveness of deep learning was successfully verified by using deep learning for cloth defect identification and detection experiments.Moreover,as the number of cloth data sets increases,there is reason to believe that deep learning will be more successful in the field of cloth defect identification and detection. |
