Global vs. Local: Component Based Learning for Classification and Image Understanding.
详细信息
文摘
In this dissertation, we describe several component based learning approaches to combining global and local information in datasets or images for classification and image understanding. Firstly, we propose the compositional sparse coding method, in which each image is assumed as a composition of several repeated patterns selected from a learned dictionary. The learning of such a dictionary of repeated patterns is unsupervised. It iterates an image encoding step and a dictionary re-learning step. Experimental results show that the proposed approach is capable of learning meaningful repeated patterns from training images and the learned repeated patterns are useful for image classification. Secondly, we propose a novel approach for data classification that combines advantages of generative model and discriminative model. We called it as the local-generative-global-discriminative model. It has two major components: a group of local generative models and a global discriminative model. The proposed approach employs a boosting framework to pursuit a group of local generative models and trains a global discriminative model as the classifier to fuse these local generative models to do the classification. The proposed approach is compared with several famous classification algorithms such as SVM and AdaBoost, and higher accuracies are obtained. Finally, we propose a distance metric learning approach for classification and dimensionality reduction. In the proposed approach, a mixture of sparse distance matrices are learned and each component in the mixture model is a neighborhood component analysis model. We test the proposed distance metric learning approach on several UCI machine learning datasets and image classification datasets, and very good results are achieved.