Accelerated Defect Detection using Memory Vector Scaling

Abstract:
Fabric defect detection plays an important role in production lines to ensure product quality and reduce defect rates. Although convolutional autoencoders have been widely used in the field of unsupervised defect detection, few researchers have previously explored the acceleration of fabric defect detection. Therefore, this paper aims to propose a model acceleration method applicable to fabric defect detection, taking the basic structure of autoencoders as a starting point. First, this paper designs a real-time and efficient autoencoder network structure using memory vector dimensionality deflation. Second, the idea of image restoration is used to introduce irregular masking in the input image during the training phase to prevent the autoencoder from replication. Then, a memory vector information fusion module with a small increase in computation is proposed to improve the expression capability of the memory vectors. Finally, an experimental comparison with the state-of-the-art algorithm is performed on a fabric dataset of three materials collected on the pipeline, and the method in this paper achieves higher performance and faster fabric defect detection.