Framework Design for Similar Video Detection: A Graph Based Video Clustering Approach

Partial duplication of existing videos is a widespread phenomenon afflicting the world. Old videos are re-circulated to fit an agenda or narrative. Such videos cause misunderstandings, often leading to clashes and violence. It is challenging for media companies to keep track of proprietary videos being trimmed and shared across various content-sharing platforms. Thus, the duplication of videos has social and economic repercussions and is a growing menace in the digital age. In this work, we propose a strategy for efficiently identifying duplicated videos. We create a system that first scans videos for important frames (called keyframes), collects and stores their hash values, and finally matches the hash of the keyframes of a new video with the existing ones in the archive to find a match. We have improved the existing methods of keyframe detection with an additional step of window selection to maintain a smooth transition between keyframes. We have used various hashing algorithms like perceptual hashing, differential hashing, and average hashing to find similarities between keyframes. Also, we have tested our pipeline on a recently published Partial Video Copy Detection (PVCD) dataset, which contains highly perturbed versions of the same video. Our method achieved an accuracy of 94% for a subset of the dataset. We finally propose a graph-based architecture to arrange the videos with similar ones clustered together. For any new video, we follow an innovative search strategy, comparing it with a representative video from each cluster and then iterating through selected clusters with high similarity scores.