Conventional image restoration techniques are based on some assumptions about a degradation process and the statistics of the additive noise. Logically, a linear model will not be able to perform restoration satisfactorily if a blurring function is strongly nonlinear. This paper aims to develop a technique for nonlinear image restoration using a machine learning approach, where no prior knowledge and assumptions about the blurring process and the additive noise are required. Although some similar learning image restoration methods exist, no one has explored the mechanism to determine why a mapping neural network can be used in modeling the degradation process, and why it works well for some images but does not behave properly for others. In this work, we try to get a better understanding about these. A generic nonlinear image restoration model is considered in this paper. Based on our previous study, a standard radial basis function (RBF) network is employed to realize the functional mapping from the degraded image space to the original image space, which is dynamically structured to ensure good generalization in restoration. The proposed adaptive RBF network is implemented in a dynamic component-based software framework, which can run in either sequential or parallel modes. Primary simulation results indicate that our proposed method perform well in restoring spatially invariant images degraded by nonlinear distortion and noise.
[1]
Uday B. Desai,et al.
Image restoration using a multilayer perceptron with a multilevel sigmoidal function
,
1993,
IEEE Trans. Signal Process..
[2]
Saleem A. Kassam,et al.
RBFN restoration of nonlinearly degraded images
,
1996,
IEEE Trans. Image Process..
[3]
Jaakko Astola,et al.
A new class of nonlinear filters-neural filters
,
1993,
IEEE Trans. Signal Process..
[4]
Tharam S. Dillon,et al.
Theoretical Foundation for Nonlinear Edge-Preserving Regularized Learning Image Restoration
,
2002,
Australian Joint Conference on Artificial Intelligence.
[5]
José M. N. Leitão,et al.
Sequential and parallel image restoration: neural network implementations
,
1994,
IEEE Trans. Image Process..
[6]
Alex Talevski,et al.
A dynamically reconfigurable component-based architecture
,
2002
.
[7]
Krzysztof J. Cios,et al.
Image Recognition Neural Network: IRNN
,
1995,
Neurocomputing.
[8]
Simon Haykin,et al.
Neural Networks: A Comprehensive Foundation
,
1998
.