Neural nets for image restoration

摘要

No imaging system in practice is perfect, in fact the recorded images are always distorted or of finite resolution. An image recording system can be modeled by a Fredholm integral equation of the first kind. An inversion of the kernel representing the system, in the presence of noise, is an ill posed problem. The direct inversion often yields an unacceptable solution. In this paper, we suggest an Artificial Neural Network (ANN) architecture to solve ill posed problems in the presence of noise. We use two types of neuron like processing units: the units that use the weighted sum and the units that use the weighted product. The weights in the model are initialized using the eigen vectors of the kernel matrix that characterizes the recording system. We assume the solution to be a sample function of a wide sense stationary process with a known auto-correlation function. As an illustration, we consider two images that are degraded by motion blur.