ABSTRACT
In this chapter, we introduce the technique of model-based neural networks, which is to be applied to our problem of adaptive regularization in image restoration. Instead of adopting the general neural network architecture introduced in Chapter 1 for our adaptive image processing applications, we propose the use of modular model-based neural network for our purpose. We used the term “modelbased” in the sense of Caelli et al. [62], where expert knowledge in the problem domain is explicitly incorporated into a neural network by restricting the domain of the network weights to a suitable subspace in which the solution of the problem resides. In this way, a single weight vector can be uniquely specified by a small number of parameters and the “curse of dimensionality” problem is partially alleviated.
