Multiset-Based Image Segmentation
Luciano da Fontoura Costa · HAL (Le Centre pour la Communication Scientifique Directe) · 2021
Given an image, the identification of its constituent regions and components represents of the most challenging tasks in image analysis and pattern recognition, commonly known as image segmentation. In the present work, we show how recently developed concepts and methods based on the real-valued Jaccard and coincidence methods can be applied in order to achieve effective image segmentation. These concepts and methods are largely related to the generalization of multisets to real, possibly negative multiplicities. The two aforementioned coincidence indices are applied to quantify, in a strict and controllable manner, the similarity between every pair of pixes, yielding a respective complex network representation with enhanced modular structure. A parameter controlling the relative influence of pairs of pixels with the same or opposite signs on the overall result allows the selection of a suitable segmentation, which is then performed by using standard community finding algorithms.