Content-based Classification Of Short Animation

With the rapid development of Internet technique, there are more and more multimedia information on Internet. Especially in the recent several years, following music and images, the digital animation became a new kind of digital media to carry information on Internet. Therefore, there is a need for a new method to classify, retrieve and filter animation. The CBIR techniques which developed rapidly in the past years may perform well for still images but may not achieve satisfactory results on dynamic animations since they are not designed for animation and could not be applied to the classification, retrieval and filtration of animation.Thus, in this paper, we managed to propose a new method for the content-based classification of short animation which based on the traditional CBIR techniques. Since many animations nowadays are usually advertisements and contain only junk information, it is valuable to propose a method to detect and filter such information, and in this paper, the classification of animation would be based on these two classes. This paper would focus on the feature extraction and analysis since the main differences between the classification of animations and images are the different method of feature extraction while the classifier does not where the input vectors come from. Firstly, this thesis introduces current research trends, CBIR system architectures and key technical knowledge, particular the methods of semantic feature extraction and analysis, and the usual way of classification. Since the particularity of the classification target in this paper, other methods of feature extraction such as character region recognition are introduced.After the feature extraction, the mutual information (MI) is employed to analyze the validity of different features, discriminative skills of are compared in this section. We also compare the differences between considering one animation whole or a serial of independent pictures and point out that the latter performs worse.Finally, SVM with RBF kernel is employed as the classifier to validate the results of feature analysis. The classification results of not only single features, but also different feature combinations are compared in this section. The conclusion of feature analysis is proved by the results of classifier and an average error of 8.28% is achieved by the optimum feature set.