ABSTRACT
For the purpose of the retrieval of impressive video scenes from lifelog videos, an efficient scene detection method is proposed in this paper. A video scene could be assumed to be impressive when it includes a person with some kinds of emotions. Therefore, the impressive scenes are detected based on facial expression recognition. The proposed recognition method is considerably efficient by introducing several types of facial features easily obtained from the positional relationships of facial feature points such as the end points of eyebrows, eyes and a mouth. By using them, multiple facial expression recognition models are constructed and integrated based on an ensemble learning approach. Additionally, an efficient emotional video scene detection method is introduced. It detects emotional scenes by finding a beginning and an ending frames of emotion expression from the recognition results of all frame images. Several emotional scenes are integrated by hierarchical clustering so that the video scenes retrieved have appropriate lengths. The detection performance of the proposed method is evaluated through an experiment using lifelog videos.
