Slide localization in video sequence by using a rapid and suitable segmentation in marginal space

Sheetal Rajgure, Vincent Oria, Pierre Gouton

Abstract: The paper is interested in indexing lecture videos for semantic search of learning material. We present a comparative study between DCT, Grayscale and Marginal space using classical k-means technique. For stable segmentation we propose to introduce an automatic threshold based on moment preservation. We discuss the suitability of each space on different images and then focus on educational video frames that are not so predominant in color and present the best transformation technique for segmentation of lecture video frame. We also present a technique to localize the slide based on some heuristics.