Multi-Modal Deep Learning based Metadata Extensions for Video Clipping

General video search and recommendation systems primarily rely on metadata and personal information. Metadata includes file names, keywords, tags, and genres, among others, and is used to describe the video's content. The video platform assesses the relevance of user search queries to the video metadata and presents search results in order of highest relevance. Recommendations are based on videos with metadata judged to be similar to the one the user is currently watching. Most platforms offer search and recommendation services by employing separate algorithms for metadata and personal information. Therefore, metadata plays a vital role in video search. Video service platforms develop various algorithms to provide users with more accurate search results and recommendations. Quantifying video similarity is essential to enhance the accuracy of search results and recommendations. Since content producers primarily provide basic metadata, it can be abused. Additionally, the resemblance between similar video segments may diminish depending on its duration. This paper proposes a metadata expansion model that utilizes object recognition and Speech-to-Text (STT) technology. The model selects key objects by analyzing the frequency of their appearance in the video, extracts audio separately, transcribes it into text, and extracts the script. Scripts are quantified by tokenizing them into words using text-mining techniques. By augmenting metadata with key objects and script tokens, various video content search and recommendation platforms are expected to deliver results closer to user search terms and recommend related content.