Fabric defect detection using deep learning: An Improved Faster R-approach
Meng An, Shiyu Wang, Liaomo Zheng, Xin-Jun Liu
Abstract
In view of the various types of fabric defects, complex textures and the lack of existing defect detection technology, this paper proposes a new fabric defect detection technology based on deep learning. Specifically, based on the Faster R-CNN network model, a deep residual network is used instead of the traditional VGG-16 for feature extraction. In order to increase detailed shallow features, a feature pyramid model of different scales is constructed as the input of the RPN network. And the number of anchors are increased to adapt to small object detection scenarios. In addition, the regularized softmax classifier is used for training to improve the network convergence ability and classification accuracy. Experimental results on fabric defect dataset show that the improved model has fast convergence speed and excellent model performance. The average precision on the defect dataset has reached 94.66%, which is 4.35% higher than the original model.