Image Inpainting

Abstract

Image inpainting is the process of filling an unknown region of an image with visual plausible information. Filling this region with information that could have been in the image. There is nothing like the 'perfect' inpainting algorithm, each method has it own advantages and drawbacks. In this thesis we want to investigate if we can improve both the speed and the visual appeal of two popular algorithms. First we look at the Bertalmio method, second at the Criminisi method. The two algorithms were implemented and analyzed by looking at the advantages and drawbacks of these methods. From these drawbacks we came up with contributions to increase the visual appeal or the speed of these algorithms.

For the Bertalmio paper three contribution are found, these are using multi-resolution images, inpainting the unknown region inwards and estimating the amount of necessary iterations of the algorithm. These contributions all increase the speed of the algorithm. Three contributions are also found for the Criminisi paper. These are local searching which increases the speed but should decrease the visual appeal, although in the experiments it increased the visual quality. Patch estimation increases the quality of the resulting images at the cost of a small speed decrease. The use of a lookup data structure significantly increases the speed of the algorithm at the cost of a small quality drop.

Observations showed that the current numerical quality estimation is not always ideal. Therefore an user study is conducted to investigate if this quantitative estimation matches the quality of the images as perceived by human beings. This user study showed that three observation were verified.

The main conclusions of the project are that contributions were found of two important inpainting algorithms and the observations from the project were verified by the user study.