Automatic Topic-Based Lecture Video Segmentation

Abstract

The recent rise in popularity of online learning has completely overhauled the way students consume informative content and how educators provide it. One of the ways teachers can provide more engaging lectures online is by segmenting recorded lectures by topic, which allows students to interact with the video and navigate it. Unfortunately, manually segmenting videos requires time and effort, adding even more to their workload. To address this issue, a system that takes any slide-based lecture videos as input and outputs a list of all the topic transitions contained in the video was developed. Previous research on this topic have placed heavy emphasis on the lecturer’s speech to segment the video by topic. However, this research investigates the use of lecture video’s presentation slides to determine topics. Topic transitions were determined using a Convolutional Neural Network - based Binary Classification Model trained on an original dataset of lecture videos collected from different educational resources. The video undergoes a series of preprocessing steps that gradually cut down the group of frames to contain only distinct slides before inputting them into the model. The classification model’s performance was satisfactory. It obtained 91% accuracy and an 80% F1 score, which is indicative of its reliability in determining whether a slide is a topic transition or not. The system developed in this research provides both lecturers and students with a method to label and segment their videos based on key topics automatically. With further evaluation, this system can potentially be proven to be convenient enough to be introduced as a new tool in the educational industry to supplement online learning.