MULTI-LEVEL VIDEO ANALYTICS FOR EFFICIENT DEDUPLICATION IN CLOUD STORAGE

Abstract

A video search query resulting in several similar videos is a common search problem users encounter these days while they expect few relevant videos matching their search context. Also, redundant/duplicate videos lead to inefficient utilization of cloud storage which is the cost-effective storage now a day. Extensive exploration of video deduplication techniques has good scope for research. Hence, the proposed multi-level video analytics aims to identify if a given input video is a possible duplicate of existing videos. Three-level similarity check namely summary_check, transcript_check and video_frame_check is proposed for duplicate determination. Summary_check level uses cosine similarity metric for determining semantic likeness of a set of documents. Transcript_check level uses text-to-speech service provider called AssemblyAI which provides the transcripts of uploaded video or audio files for comparison. Video_frame_check level based on deep learning algorithm compares videos based on the search relevant objects present in the video. The threshold defined for each level decides the exactness of duplicates and metadata generated in each level is stored in a database for future use. The input video is stored in cloud storage if found to be unique. The results obtained prove to be significantly better compared to existing approaches of deduplication check.