Research On Affective-Oriented Moive Background Music Classification

With the rapid development of the movie, the popularity of the computer and the network, movie media data grows at an amazing speed. Due to the vast amount movie data, and difficulty in maintaining consistency of movie labels like movie style or affection style by artificial annotation, the traditional movie management based on artificial annotation can't meet the demand of retrieval. This new demand promotes the development of audience-oriented affection content analysis. Movie background music plays an irreplaceable role of strengthening film affection, heightening the drama and render atmosphere. If we can classify movie background music by affection, there will remarkably improve movie affective content analysis accuracy undoubtedly.So far, there are few researches on movie background music, but many researches on affective-oritented music classification. Based on systematical summarization and analysis of current music affection classification methods, movie background music affection feature vector and classifier are proposed. Movie background music affection vector is constituted by bar-long rhythm patterns, bar-long the bassline patterns, Mel frequency cepstrum coefficients and interval features extracted from music audio signal. Compared with other features, rhythm pattern and bassline pattern features are able to demonstrate rhythm structure over movie background music clip. Probability latent semantic analysis, initially used for mining latent semantic of synonym and homographs in text, is used to classify the movie background music into excitement, temsion, relaxation and sadness.Experimental results show that the movie background music affection feature vector and the PLSA classifer effectively improved the accuracy of classes than the existed classification methods in literature. But the discernibility ability of the proposed feature vector needs improving, especially at valence in V-A affective space. Features with better discernibility should be extracted in future work.