Deep Learning-Based Automatic Classification of Slide Frames and Presenter Frames

Abstract

Online learning has become very popular among the current generation. In many of the online learning contents, the only shared content is only the video which has the presenter making the presentation interspersed with slides from the presentation. Presentation slides or lesson notes are sometimes not shared or not available. Our current work explores automatic lesson notes creation from education lecture videos. The final goal is to make presentation slides of the lecture content from the video. As a first step in this regard, we currently explore automatic classification of video frames as either being slide frames or presenter frames. We have used a pre-trained deep neural network—ResNet trained on Imagenet. Our data is frames generated from cs231n Stanford course. We had retrained the fully connected layers of our model on our data. Our experiment results show that we are able to classify slide frames and presenter frames with 98.8% accuracy with as less as 20% of our datasets as training. Our approach is promising in getting to the final goal of automatically creating presentation slides from lecture videos.