Yarn-dyed Shirt Piece Defect Detection Based on U-shaped Swin Transformer Auto-encoder

Abstract

Automatic defect detection is an essential and challenging problem in the yarn-dyed weaving production process, this paper proposed a novel U-shaped Swin Transformer auto-encoder reconstructed model for yarn-dyed shirt piece defect detection. This method uses the model of Transformer to extract the global features of the image better and reconstruct it more accurately, which solves the problems of scarce and unbalanced type of defect samples and high cost in actual production. Firstly, for a certain pattern, using defect-free samples adding Gaussian noise to train the reconstruction model. Then, the image to be tested is input into the Transformer model to obtain the corresponding output image. Subsequently, the residual image is calculated by subtracting the input image and its corresponding output image. Finally, the defect localization can be achieved through thresholding and morphological operation. The experiment result verifies the effectiveness of the proposed method on various types of yarn-dyed shirt pieces.