Scene Completion Using Top-1 Similar Image

Abstract

Scene Completion is an interesting Image Processing problem that has recently been studied in the context of data, i.e. by using large repositories of data. The idea is to extract the most suitable image from a data repository that could later be used in constructing the missing part of the input image. One of the requirements for such a data intensive approach is that the completion has to be completed without human intervention. This is rather challenging as it may not be clear that what could be the most suitable image in the data repository for the completion purpose. Thus, whilst it is important that the repository should be of reasonable size so that a matching image could be found, it is even more important that we are able to retrieve that top-1 image that could be the best candidate for scene completion in an automated fashion. This is the focus of this work, i.e. we propose a methodology for finding the top-1 image in a data repository that could be the best candidate for scene completion. Then we give an algorithm for scene completion that works as follows. In the first step, matching images with the query image are retrieved that could be the candidates for the completion. We consider three features for the purpose namely Gist, Texture and Colour. Gist feature represents the shape and structure of the image. Texture helps when segments from different images are being combined together. Similarly, the Colour feature helps in finding and harmonising the colour difference between the input image and the candidate completion image. Next, we propose a ranking scheme that retrieves top-1 most suitable candidate image. The ranking scheme is able to work with any number of features computed and also satisfies the value-Invariance property.