Defect Detection of Yarn-dyed Shirts Based on Denoising Convolutional Self-encoder

Abstract

A defect detection algorithm for dyed shirts with supervised framework relies on a large number of labeling samples and high modeling cost. This paper proposes an automatic detection and location method for color fabric defects based on unsupervised denoising convolution self-encoder. Firstly, a shirt image data set containing 66 kinds of yarn-dyed patterns and a total of 11900 pieces was constructed. Then, Gaussian noise was added to the defect-free samples, and a depth-deconvolution convolution self-encoder was used to construct the image of the yarn-dyed shirt piece. The denoising reconstruction model performs reconstructive repair on the noise interference; Then, the residual of the image is tested and the reconstructed image is calculated, and the slice defect region is detected and located using a mathematical morphology algorithm. The experimental results show that the image reconstruction model and residual image analysis algorithm based on denoising convolution self-encoder can effectively detect and locate the defective area of the dyed shirt piece without relying on the label sample.