University of Houston

Lecture videos have been commonly used to supplement in-class teaching and for distance learning. They are versatile learning resources; but the usage of long lecture videos is limited to video format which makes a challenge to quickly access the content of interest. Indexed Captioned Searchable (ICS) project at UH aims to understand the perceived value of these lecture videos and increase the accessibility to the content of interest by its features. Indexing adds logical index points, each in the form of a snapshot representing a video segment that can be accessed directly. Search enables identification of video segments that match a keyword provided by the user. In this work, we present a text based video indexing algorithm using the text extracted by Optical Recognition Tools (OCR). We also present image enhancements -text segmentation and inversion-, to increase the detection accuracy of OCR tools. Lecture video indexing includes identifying all Transition Points (TP), i.e., places where the scene on the video changes significantly. Text of each TP is extracted by OCR tool and comparison done between the texts of adjacent TPs. Texts are represented by word frequency vectors and cosine similarity metric is used for comparison in text based indexing. To compare and evaluate different indexing algorithms, we have provided a new framework. Creating the ground truth data for ambiguous index points and evaluating based on the new accuracy metric is presented in this framework. Preliminary results on selected diverse 25 videos show that text based indexing algorithm provides more accuracy than image based indexing algorithms which based on scene changes. We propose to leveraging the current approach by additional experiments in more videos and using the semantic similarity of words for text comparison. We also propose using ground truth data for new machine learning based lecture video indexing algorithm.