Yarn-Dyed Shirt cut Pieces Defect Detection Using Attention Vector Quantized-Variational Autoencoder

Abstract

For yarn-dyed shirt cut defects detection problems in production process, this paper proposes a yarn-dyed shirt cut defects detection method based on attention vector-quantized variational autoencoder reconstructed model and residual analysis. To solve the actual problem that the defect sample quantity scarce, defect categories imbalances, high cost and poor generalization ability of artificial design defect features. Firstly, for a certain yarn-dyed shirt cut, salt and pepper noise is artificially added to the defect-free samples to construct a training data set, and then a reconstruction model based on the attention vector quantized variational autoencoder is established and trained. Secondly, a residual map between the original image and the correspondingly reconstructed image is calculated. Finally, the defective area could be detected and located by thresholding and opening operation processing. Experimental results on several yarn-dyed shirt cut pieces data sets show that the proposed method can effectively reconstruct the yarn-dyed shirt cut pieces, detect and locate the defect area of yarn-dyed shirt cut pieces quickly.