Content Based Video Retrieval for Detecting Videos with Irrelevant Content for Children

Abstract

With advancements in networking technologies, content produced and distributed over the Internet is exponentially increasing. This imposes the threat of distribution of offensive adult content freely and largely, urging mechanisms to control access by minor aged users to such content. Manual retrieval and indexing of materials for this matter is cumbersome and impossible for large data repositories, hence require automatic means for effectively achieving the foresaid operations. Most of the automatic approaches used in existing systems are based on filtering by explicit metadata such as tags and textual content in urls. This project aims at using content-based video retrieval approaches to detect videos with offensive adult content, specifically those that contain nudity, which are not suitable for children. We propose a solution based on the Bag of Visual Words (BoVW) algorithm to solve the aforementioned issue. Our solution initially obtains a summary of video as keyframes and apply BoVW algorithm in order to classify keyframes for the purpose of indicating the presence of obscene content. We propose a novel algorithm to summarize the video by extracting keyframes that mark video shot boundaries. This algorithm computes an adaptive threshold derived for a moving window of frames and uses that threshold to filter redundant video frames. The proposed algorithm reduces the redundancy up to 99% in videos. The proposed solution for recognizing obscene content in videos is based on spatial information in single video frames. Despite the ignorance of high-level features in temporal domain, we prove a recognition rate of 85% with spatial information alone. Further, we have proved the irrelevance of color information to detect nudity in videos when using BoVW algorithm. We also have proposed future work that can be carried out to extend the proposed work to enhance the recognition accuracy.