Learning-Based Image Registration Approach Using View Synthesis Framework (LIRVS)

Abstract

Image registration is a fundamental preprocessing step in varied image processing and computer vision applications, which rely on accurate spatial transformation between source and the reference image. Registration using iterative view synthesis algorithm is experimentally shown to solve a wide range of registration problems, albeit at the cost of additional memory and time to be spent on the generation of views and feature extraction across all the views. Hence, we have approached the problem by building a decision-maker model which could predetermine the possibility of registering the given input image pairs and also predict the iteration at which the image pair will be registered. The proposed approach incorporates a decision-maker (trained classifier model) into the registration pipeline. In order to ensure that the gain in time is considerable, the classifier model is designed using the registration parameters obtained from reference and source image. The trained classifier model can predetermine the possibility of registering the input image pairs and also the minimum number of synthetic views or iteration necessary to register the input image pair. Hence, for the images that have been registered, an additional time required for the proposed approach is tolerable. However, for the images that are not registered, the gain in time because of classifier model is extremely significant.