Machine Learning: Any Image Reconstruction Algorithm Can Learn by Itself

Abstract

In the traditional machine learning sense, a neural network structure with single or multiple layers is required and the weights at the neurons are learned from training sets. The main application of machine learning is pattern classification. The author’s main research area is image reconstruction. Any image reconstruction or image processing algorithm has some parameters to be adjusted according to the tasks. For example in Bayesian algorithms, the weighting factor for the Bayesian influence and some thresholds for non-linear constraints are tedious and time-consuming to determine. In this preliminary research, we argue that any image reconstruction algorithm or any image processing algorithm can learn to find its optimal parameters by using the training sets. The neural network structure is not necessary.