Learning video editing profiles

Abstract

In this project, we study three related topics for video mining, video learning and video recognition. In video mining, the attribute relevancy analysis is applied to evaluate the usefulness of different editing parameters. Two different algorithms are implemented, one for the sequential pattern mining and one for the periodic pattern mining, to describe the common editing patterns found in different video generes. In addition to mining, we present a methodology for video learning and video recognition through the use of Hidden Markov Model (HMM). We analyze the shot parameters in videos of different genres and model each video genre as a HMM with shot parameters as their observation vectors. The most likely underlying state sequence that give rise to the observation can be derived using Viterbi Alignment and Baum-Welch Algorithm. Video learning is to classify different video genres according to the video parameters we define. Each HMM is trained with a sufficient number of representative examples. Video Recognition is to recognize different video genres correctly. A complete set of HMMs is used to recognize videos of different genres.