Neural Network Segmentation of Video Via Time Series Analysis

摘要

Semantic video retrieval which deals with unstructured information traditionally relies on shot boundarydetection and key frames extraction. For content interpretation and for similarity matching between shots, videosegmentation, i.e. detection of similarity-based events, are closely related with multidimensional time seriesrepresenting video in a feature space. Since video has a high degree of frame-to-frame-correlation, semantic gapsearch is quite difficult as it requires high-level knowledge and often depends on a particular domain application.Based on principal components analysis a method of video disharmony authentication has been proposed.Regions features induced by traditional frame segmentations have been used to detect video shots. Results ofexperiments with endoscopic video are discussed.