Parameter-free hierarchical image segmentation

Images typically have many levels of detail and the suitability of a segmentation depends on application requirements. Thus, it is imperative that the user/application be given the option to select the ‘optimal’ segmentation that captures the desired level of detail from a set of segmentations. This paper presents a hierarchical image segmentation algorithm that offers this option using the concept of minimum spanning trees. It converts an input image into a tree structure from which a hierarchy of segmentations is obtained through a process of merging. No parameters are used in this process and thus the proposed algorithm can be used on any segmentation dataset as is. The levels are calculated in one pass of the minimum spanning tree and as such, no iterative merging is required. This provides the user with a quick way of segment visualisation. Evaluation results on two popular segmentation datasets show that the algorithm provides competitive results in comparison to other segmentation algorithms.