When: Oct. 10 2019, 17:00-18:15
Where: Bldg. 110, Room 103

Speaker: Gyeongha Hwang (Youngnam University)

Title: GAN-Adaptive Auxiliary Information-Free Deep Learning in Solving Linear Inverse Problems

Abstract: In this talk, we address the linear inverse problem with the goal of recovering the original data from the observed data, which may be contaminated with noise. As a fundamental problem in diverse applications, this has been studied extensively in past several decades from diverse points of view. One of the most popular methods to get a better solution is to exploit a hand-designed signal prior. Although it is generic and applicable to various problems, it does not guarantee high reconstruction quality. A new direction and trend is the use of deep learning to learn the map ping between the original image and the observed image. While deep learning achieves state-of-the-art performance, it exhibits a weakness in that different problems require different networks. Recently, a new paradigm in the middle of the two aforementioned methods was suggested by Chang {it et al.}. They proposed a general framework to train a single deep neural networ k using synthesized inputs, which are generated by an auxiliary noise function. Their method is capable of solving arbitrary linear inverse problems. Despite its flexibility and superior performance, the key to their method is how to choose the auxiliary noise function such that the synthesized inputs in training are as consistent with the inputs in testing as possible. Nevertheless, such a choice is difficult and affects reconstruction quality since various inputs in testing depend on different linear inverse problems. We propose a new method to solve the inverse problem via deep learning, where only Generativ e Adversarial Networks (GAN) instead of any auxiliary functions are exploited. More precisely, we first replace the whole training process by GAN. Then, through a newly designed signal prior, we integrate the outcome of GAN into Alternating Direction Method of Multipliers (ADMM) to get the recovered result. Furthermore, our method is extended to solve linear inverse problems with large but structured noise. Empirical results demonstrate that our method outperforms the previous works.
List
  Edit