Automatic capsule preparation for lecture video

In this paper, we propose a visual quality and content based approach for creating a lecture video capsule which essentially contains the highlights of the original lecture video. We first segment and recognize the activities in the instructional video using a hidden Markov model(HMM). The activities are classified into three categories : talking head, writing hand and slide show. Then a no-reference objective quality assessment of the non-content frames, i.e., talking head frames, are performed to detect a few high quality frames for highlight creation. A new method of defining the visual quality of content frames, i.e., writing hand and slide show frames which are very crucial in a lecture video, is proposed in this paper. Statistical parameters of the intensity histogram along with the horizontal projection profile(HPP) of the the content frames are used to derive an objective quality measure of these frames. This allows us to extract the high quality content frames, which are `well-written'. Finally on media re-creation to produce the video capsule, we select the video segments of a few seconds duration around these good quality frames along with the audio and merge with suitable proportion of the detected classes of instructional activity. This technique accounts for a very compact representation of the whole lecture video which finds a very potential application of video previews.