GENERATION OF TEXTILE PATTERNS THROUGH GENERATIVE ADVERSARIAL NETWORKS

Abstract

In recent years, the generative adversarial networks (GANs) are used to generate photorealis tic images that apparently look real, but are generated by a deep learning algorithm. However, GANs are known to be quite unstable as sometimes they show inconsistent results. In 2015, Alec Radford and Luke Metz proposed the deep convolutional generative adversarial net work (DCGAN) that uses a specific set of modifications made to the original GAN in order to improve its training and stability. GAN generates a new image, which mimics the real image using already available data. The proposed system deals with the generation of textile patterns on the industrial scale with the assistance of DCGAN. We use the convolutional instead of pooling layers while modifying the activation functions, as proposed by Alec Radford and Luke Metz. The results demonstrate that the presented model is able to generate basic patterns from the seed images, despite the critically unlabeled data and computation limitations