Exemplar-Based Inpainting Using Shift-Map and Roof Duality
Fredrik J. E. Svensson · 2013
To inpaint a digital image is to fill in unwanted pixels in an image with other, plausible, content. One use of this is to remove unwanted objects in a photograph. A user marks the unwanted objects, and then the objects should automatically disappear. There are many ways to inpaint images, and one of them is so called exemplar-based. This means that other parts of the image is copied into the area that should be replaced. Every pixel marked by the user is mapped to another pixel in the image, from which it copies its content. This thesis investigates both if the optimization method can be improved and if the model describing the image can be improved compared to the state-of-the-art. The problem is modeled as a function that is to be minimized. It is shown that the minimum value is lowered by applying so called roof duality, which is a discrete optimization method. Also, it is demonstrated that modifications to the model can improve the qualitative result in many ways. Those aim at removing several undesired phenomena without worsening any other results, for example badly mapped high frequency content. The modications are tested and compared on real images with good results. (Less)