Mining Sequential Patterns for Image Classification in Ubiquitous Multimedia Systems

Image classification is an important technique for effective content-based multimedia retrieval. Many classifiers have been proposed while frequent patterns based approaches have received many attentions in recent years. In this paper, we proposed an image classification approach utilizing sequential patterns discovered from distinct classes. The image is segmented and low-level features are extracted as a sequence of feature-sets. Sequence-rules are collected from each class and conflict rules are resolved by rule pruning. Useful rules are then selected to form the prediction rules. Four prediction heuristics are provided to raise the prediction accuracy of the image classifier. Experimental results using synthetic datasets show that the proposed algorithm may reach an accuracy of 78%. Effective image classification thus can be achieved by mining sequential patterns for the construction of the rule-based sequence classifier.