Hierarchical Mutual Nearest Neighbour Image Segmentation

This paper presents a hierarchical image segmentation algorithm based on the principle of mutual nearest neighbours. Image segmentation remains a great challenge in the computer vision community. To solve this problem, various algorithms have been proposed in the literature. However, most of these algorithms depend heavily on thresholds or parameter settings. Furthermore, the majority of them do not recognise the hierarchical nature of the problem. In particular, there might not be a single best segmentation for an image as the level of detail that should be present in a segmentation will depend on the purpose for which that segmentation will be used. Many algorithms might provide good results for a specific application, or in detecting a certain level of detail in an image, but may fail when they are applied to different types of images at a different level of detail. The method proposed in this paper generates a hierarchy of segmentations that retain different levels of detail. Thus, depending on the application, the segmentation that provides the required level of detail can be selected. Utilisation of only one meta-parameter in the process makes it applicable to any dataset as is. It can also be easily generalised to segment different types of images including 3D and multispectral images. Evaluation on the Berkeley BSD500 dataset and comparison with existing hierarchical segmentation algorithms provide superior results.