Lecture Video Summarization Using Deep Learning

In today’s rapidly evolving digital era, educational institutions and professors are increasingly sharing video lectures online to support students. While these lectures effectively teach topics from scratch, they pose challenges for quick revisions. Viewers struggle to control the pace, often interrupting playback to navigate the content. Additionally, finding specific information within the video and skimming the unstructured transcript for relevant content can be cumbersome, hindering efficient learning. The objective of this paper is to provide a framework for summarizing video lectures to facilitate quick revisions. We further aim to help students with topic-wise preparation by summarizing specific lecture sections relevant to them, saving them the extra effort of manually searching through the entire lectures. We employ shot detection, speech transcription, transcript summarization, and ontology tree to generate topic-wise slideshow summaries. The summarization models: BART, T5, GPT3, Extractive, and Seq2Seq are evaluated on the VT-Ssum dataset, and their ROUGE scores are compared to select the best model.