Institute of Information Theory and Automation

Convolutional neural networks in digital image restoration

Project leader: Mgr. Zuzana Bílková
Department: ZOI
Supported by (ID): GAUK, No. 1583117
Duration: 2017 - 2019
Publications at UTIA: list

Abstract:

Digital image acquisition is often accompanied with its degradation by noise, blur (out-of-focus, motion etc.), compression, etc. In many cases, the degradation process can be modeled by a linear relation g=Hu+n where g denotes the acquired image, u the original image, H the degradation operator, and n random noise. The goal of image reconstruction is to recover the original image based on the observed image. The recent boom of convolutional neural networks (deep learning) has to some extent improved results for a limited class of simple reconstruction problems. We plan to exploit convolutional neural networks to improve results, enlarge the family of problems and investigate applicability of generative neural networks to represent prior information. We believe the latter has a potential to overcome the limitations of existing methods.

Project team:
Responsible for information: ZOI
Last modification: 14.03.2017
Institute of Information Theory and Automation