Variational image fusion

David Hafner

Abstract: The main goal of this work is the fusion of multiple images to a single composite that offers more information than the individual input images. We approach those fusion tasks within a variational framework. First, we present iterative schemes that are well-suited for such variational problems and related tasks. They lead to efficient algorithms that are simple to implement and well-parallelisable. Next, we design a general fusion technique that aims for an image with optimal local contrast. This is the key for a versatile method that performs well in many application areas such as multispectral imaging, decolourisation, and exposure fusion. To handle motion within an exposure set, we present the following two-step approach: First, we introduce the complete rank transform to design an optic flow approach that is robust against severe illumination changes. Second, we eliminate remaining misalignments by means of brightness transfer functions that relate the brightness values between frames. Additional knowledge about the exposure set enables us to propose the first fully coupled method that jointly computes an aligned high dynamic range image and dense displacement fields. Finally, we present a technique that infers depth information from differently focused images. In this context, we additionally introduce a novel second order regulariser that adapts to the image structure in an anisotropic way.