Effective Video Retrieval System using Adaptive Shot detection and feature vector algorithm
Snehal Patel, Vivek Deodeshmukh · 2017
In current scenario roughly 100 million hours of videos are being uploaded on the internet (Viz. YouTube, Netflix, Dailymotion, Vimeo, Veoh, Metacafe, etc.) It becomes very difficult to extract a desired related video from such huge data set. Semantic/context based matching is fast but highly dependent on correct tag assigned to a video. On the other hand it is really difficult to apply and context based search on the video, because of large number of frames involved in a video. We have developed a novel video retrieval system, which can extract the desired videos from the large video data set. Algorithm consists of content based adaptive shot detection and feature vector extraction for each data set video. User can search any video in the data set by just giving a similar image as system input. We have achieved accuracy of 97.2% with blind trials and 100% accuracy in case of data base frame based search. Our algorithm also reduces the search time to roughly 0.3ms/video which is 100 times faster than reported search algorithms in literature. We expect our algorithm would fit in as an alternative to current video search algorithms and it can also act as image based video search along with existing technology of text based video search (e.g. YouTube).