Probabilistic Inferencing and Markov Random Field Modelling for Image Restoration

Abstract

Image restoration is a key research topic in the area of Computer Vision. In today's world computers are being trained to be made as intelligent as humans, so that they can perceive deformed/corrupted images as efficiently as humans perceive them. In the dissertation, we will be discussing about types of Image Restoration namely Image Inpainting and Image Denoising. There are various techniques developed by researchers for Image restoration. We will be focusing mainly on Probabilistic Inferencing and Markov Random Field Modelling. Probabilistic Inferencing as the name suggests tries to infer the true image from the corrupted image by modelling probability densities. It mainly focuses on developing a model where at each step we can deduce how probable it is to obtain a completely restored image from a corrupted Image. A model called Field of Experts for probabilistic inferencing has been discussed to achieve Image restoration, The results of this technique are good in many cases but they tend to fail in cases where edge preservation is required. To meet with the failures of the above mentioned approach we explore the approach of Belief propagation coupled with Gaussian Graphical modelsin Markov Random Field modelling. Wherein a model has been developed to perform Image restoration while tapping the Markovian property that a variable is dependent only on its neighbourhood variables.On analysis it was found that this approach gave better results for Image Inpainting and helped in preserving the edges and contours of objects in images while Inpainting.