Video saliency computational model and its application in video advertising

This paper proposes a novel video saliency computational model based on frequency domain analysis, which is simple and efficient. First, the original video is parsed to a series of consecutive video scene images. Then, the space domain saliency and time domain saliency of the video scene images are computed, respectively. In each domain, the saliency of each frame is estimated by using Renyi entropy. Finally, both space domain and time domain results are fused to obtain the overall saliency. Application results in video advertising system and comparison with state-of-the-art approach demonstrate the effectiveness of the proposed model.