Geodesic distance based interactive image segmentation

Interactive image segmentation or foreground extraction is an important component of image editing tools. Studies on separating foreground from background quickly, practically and with little error has grown enormously over the last decade. Interactive image segmentation has usage in photo mounting, (medical) image processing etc. In general, interactive segmentation by definition is classifying pixels as either foreground or background which are not indicated by users using the information of labeled pixels. Labeling both foreground and background pixels separately is not the only way of expressing object of interest. Other than that, there are ways like drawing a bounding box on object or scribbling object boundaries. However, proposed methods differ according to the user interaction. For example, a method working for the former interaction type will not work for the bounding box where only background pixels are indicated. That is why only methods using the former and the last interaction type is considered in this work. In this thesis, I worked on estimating the label of unlabeled pixels by using foreground and background information obtained from user and developed a new method. The proposed method first does over segmentation on image and groups similar pixels. Thus, pixel count is greatly reduced. Then, probability of pixels being belonged to foreground or background are calculated by taking into account the spatial location of pixels to the scribble locations. Later, by using these probabilities geodesic distances to user scribbles are calculated. Pixels having distance below a certain value are labeled. Finally, remaining unlabeled pixels are labeled by calculating probabilities using the spatially closest labeled pixel information. Obtained results are compared with the methods using the same marking technique quantitatively on a publicly available dataset.