University of Houston

Lecture videos serve as an extremely useful learning resource and considered equally important as textbooks among the students of University of Houston. However, due to the continuous nature, their usability is highly limited by the inability to directly access the topics of interest within the video. Providing index that represents the topics within the video can highly improve its accessibility. Manual indexing by identifying the topics within the video is a time consuming task. Video segmentation based on identifying the scene changes is a possible solution, however the result may not match the topic transitions. On the other hand, the text within the lecture video represents the topics within the video. This thesis proposes an automated text based approach for indexing of lecture videos that can provide topic-based segmentation, thereby enhancing its accessibility.

The proposed method relies on the similarity of the text segments; therefore, the choice of similarity metric is crucial for determining the topic indices. Text similarity metric based on term-frequency vector model is highly effective in information retrieval domain for the indexing of documents and their search from repositories, the context similar to lecture video indexing. Therefore, term frequency based similarity metric forms the basis of the indexing algorithm.

The video is split into smaller segments at slide transition points by detecting the scene changes, which is determined by the image difference between consecutive frames of the video. Optical character recognition (OCR) technology extracts the text from each segment for similarity analysis. The indexing algorithm determines the topic changes within the video segments based on the text similarity.

The evaluation of the algorithm involved the selection of twenty-three videos from diverse subjects such as Computer Science, Biology and Chemistry. Based on the initial result, algorithm was fine tuned to compensate for the drawbacks. Text based indexing produced 13% improvement over non-text based methods and the results closely matched the absolute ground truth. The results also revealed the limitations of text based indexing. The author suggests future studies on hybrid methods by combining text and image as well as considering the semantics to further improve the indexing accuracy.