Improving the Performance of Video Summarization Using Deep Learning

Abstract

There has been an exponential growth in the amount of video data in the past few years. Processing massive amount of data requires resources such as time, storage, processing power, to name a few. Video summarization is to generate a summary of a long video document by selecting the most informative or interesting materials for potential users. Video summarization provides a clear analysis of long videos by removing redundant keyframes and extracting only informative keyframes. The report describes various methods used for video summarization in various application domain that include egocentric, sports, multiview, movies, news, surveillance, generic, educational, medical, and robotics. Highlights of cricket match includes important events of a cricket match such as six, out, no ball, wide, to name a few. Umpire take decision based on these events and signals using different hand actions for different events. We implemented a system to generate automatic highlight of long cricket video. The proposed approach recognizes the umpire signals by using umpire hand poses to generate cricket match highlight. In the proposed system, VGG19 and Inception V3 are used for feature extraction and SVM is trained on extracted features. SNOW dataset is used for performance evaluation. A linear SVM classifier was used to obtain the results. The SVM classifier with features extracted from the FC1 layer of the VGG19 network produced the best classification results. Built classifier is used to implement a system to generate highlight of cricket video. Video having various events of cricket is used as input for built classifier and number of events classified true are counted to calculate true positive rate. The preliminary results shows that the proposed system provides higher true positive rate compare to base system.