Fast Detection of Fabric Defects based on Neural Networks

Abstract

Anomaly detection is an important research topic in artificial intelligent studies. Among anomaly detection applications, fabric defect detections obtain lots of research interests due to its industrial potential. This study presents an efficient method to detect fabric defect regions by the Siamese network for greatly reducing the training time by only using limited training data. The model identifies texture features by using some normal and defect images. Defect regions can be detected through overlapped blocks identification and the block size determines the precisions of detection correctness and locality. At last, the proposed structure is compared with the conventional Bergmann’s autoencoder, the Alexnet-based autoencoder, and the VGG16-based autoencoder models. Experimental results show that the proposed structure requires limited training time comparing with autoencoder models and exhibits good recognition rate.