Yarn-dyed Fabric Defect Detection using U-shaped De-noising Convolutional Auto-Encoder

Abstract

Practical factors such as high labor cost of labelling defect samples and scarcity of defect samples make it difficult for supervised machine learning models to solve the problem of yarn-dyed fabric defect detection. To solve this problem, this paper proposes an unsupervised yarn-dyed fabric defect detection method based on U-shaped de-noising convolutional auto-encoder (UDCAE). Firstly, for tested samples of yarn-dyed fabric, the training dataset was constructed by collecting the non-defect yarn-dyed fabric samples. Then, the non-defect dataset is utilized to model and train the proposed UDCAE model. Finally, the defective area can be quickly detected by calculating the residual between the original tested yarn-dyed fabric image and its reconstructed item correspondingly. The experiment results show that the proposed method can accurately detect defects of yarn-dyed fabrics with different patterns.