ENDOSCOPY VIDEO SUMMARIZATION BASED ON MULTI-MODAL DESCRIPTORS AND POSSIBILISTIC UNSUPERVISED LEARNING AND FEATURE SUBSET WEIGHTING

摘要

The spread of capsule endoscopy systems has proved to be inherently constrained by the tedious diagnosis process when the physician has to review thousands of endoscopy video frames in order to detect pathology symptoms. In this paper, we propose a novel endoscopy video summarization approach based on possibilistic clustering and feature weighting algorithm. The algorithm generates possibilistic membership that represents the degree of typicality of the video frames, and that is used to identify and discard noise frames. The robustness to irrelevant features is achieved by learning optimal relevance weight for each feature subset within each cluster. We extend the proposed algorithm to find the optimal number of clusters in an unsupervised and efficient way by exploiting some properties of the possibilistic membership function. The system demonstrated promising performance in extensive testing on real-world datasets associated with the difficult problem of endoscopy video summarization. The endoscopy video collection was acquired on four patients at different geographic locations. It includes more than 90k video frames.