Restoration of degraded images with neural networks

• Published in: Systems and computers in Japan 1998-07, Vol.29 (8), p.57-67
• Main Authors: Seto, Atsushi, Ohtsuka, Kazuhiro, Ishii, Rokuya
• Format: Article
• Language: English
• Subjects: edge preservation; image restoration; neural network; noise removal; ringing control
• ISSN: 0882-1666; 1520-684X
• DOI: 10.1002/(SICI)1520-684X(199807)29:8<57::AID-SCJ7>3.0.CO;2-J

Image restoration is an important topic in the area of image processing. Standard normalization methods are well known in which this problem is treated as the problem of optimization of the evaluation function. In this paper, we treat images degraded by a given shift‐invariant blur function and additive Gaussian noise, and a method of image restoration is proposed which uses a new type of neural network. We formulate the restoration problem as the problem of minimization of a nonquadratic evaluation function which includes both the ringing effect and the smoothing conditions which control additive noise. To minimize this evaluation function, we propose a new type of neural network which is composed of a Hopfield‐type neural network (HNN) and layered neural network. The edges of the image are detected by the layered neural network, and using its output, the coupling coefficients of the HNN are dynamically controlled. Using the energy minimization principle of the HNN, the nonquadratic evaluation function is minimized. From the computer simulation results, we observed an SNR improvement on the order of a few tens of decibels, as compared with conventional methods. © 1998 Scripta Technica, Syst Comp Jpn, 29(8): 57–67, 1998