Generative Adversarial Network for Generating Low-rank Images

Low-rank texture structure is an important geometric structure in image processing. By extracting low-rank textures, images with various interferences can be rectified effectively. To solve the problem of low rank image correction with various interferences, this paper proposes to use the generation framework to alleviate poor correction results on the region without obvious low-rank properties. And a low-rank texture generative adversarial network (LR-GAN) is proposed using an unsupervised image-to-image network. Firstly, by using transform invariant low-rank textures (TILT) to guide the discriminator in the LR-GAN, the whole network can not only achieve the effect of unsupervised learning but also learn a structured low rank representation on both generation network and discrimination network. Secondly, considering that the low-rank constraint is difficult to optimize (NP-hard problem) in the loss function, we introduce a layer of the low-rank gradient filters to approach the optimal low-rank solution after many iterations guided by TILT. We evaluate the LR-GAN network on three public datasets: MNIST, SVHN and FG-NET, and verify the quality of generative low-rank images by using a classification network. Experimental results demonstrate that the proposed method is effective in both generative quality and recognition accuracy.