Optical flow estimation is a very important topic in the field of computer vision. It concerns the estimation of the movement happening in between two or more frames of a given image sequence or scene. Traditionally, variational optical flow approaches were very successful by formulating the problem as an optimization problem. More recently, with the advent of neural networks, deep learning approaches now usually outperform purely variational methods. Nevertheless, the variational concepts are also used for flow refinement in addition to the new approaches. One big problem in flow estimation are occlusions, when pixels that are visible in the first image are occluded in the second image. This thesis tries to address this problem by extending an existing variational refinement method by Maurer et al. with additional occlusion handling by utilizing external occlusion masks. These masks mark the pixels in the image either as occlusions or not. This makes it possible to address them specifically by either smoothing them or not refining them at all. Experiments with flows from the ProFlow method and the state-of-the-art deep learning method RAFT show promising results by further improving their flows. ProFlow could be improved with the smoothing approach, while RAFT showed improvement when using the area-specific refinement.