Video Segmentation And Summarization Based On Features

In the era of “big data”, revolutionary changes are taking place in digital media and related academic areas, bringing many new challenges for digital image processing. In this thesis, we focus on three major problems:? How to effectively and efficiently fuse multiple low-level features, in order to gen-erate semantic understanding of images and videos for efficient storage and index;? How to take advantages of large-scale unlabeled data and benefit feature selection for big and redundant visual features, hence removing unrelated features, reducing storage space, and improving classifiers' accuracy;? How to naturally summarize long videos for quick skimming which tries to solve “semantic gap” from another perspective.We will try to solve the above problem with the following steps:Firstly, we propose a video segmentation method via adaptive higher-order CRF with windowed dynamics, which is a general and effective resolution for moving object segmentation.Secondly, we propose a semi-supervised feature selection method for web image classication, which takes advantages of both 2,1-norm based embedded feature selection and existing large-scale unlabeled data. Experimental results on web image classification demonstrate that our method performs well with limited label information.Thirdly, as the “semantic gap” is inevitable, we propose to introduce one-class SVM to multiple low-level visual features, hence developing new algorithms towards more accurate,more effective and more representative video summarization.To sum up, we try to solve the “semantic gap” problem via taking full advantages of low-level visual features from the above three perspectives. Experimental results on realworld data demonstrate the effectiveness of our methods.