Generating Low-Rank Textures via Generative Adversarial Network

Abstract

Achieving structured low-rank representation from the original image is a challenging and significant task, owing to the capacity of the low-rank structure in expressing structured information from the real world. It is noteworthy that, most of the existing methods to obtain the low-rank textures, treat this issue as a “transformational problem”, which lead to the poor quality of the images with complex backgrounds. In order to jump out of this interference, we try to explore this issue as a “generative problem” and propose the Low-rank texture Generative Adversarial Network (LR-GAN) using an unsupervised image-to-image network. Our method generates the high-quality low-rank texture gradually from the low-rank constraint after many iterations of training. Considering that the low-rank constraint is difficult to optimize (NP-hard problem) in the loss function, we introduce the layer of the low-rank gradient filter to approach the optimal low-rank solution. Experimental results demonstrate that the proposed method is effective on both synthetic and real world images.