Discrimination Enhancement For Image Classification

Intelligent image understanding and analysis are hot and difficult research topics in the multimedia community. Image classification is the most basic and important aspect for image understanding and analysis. Image classification task is about to di-vide images into semantic categories. First, image classification uses computer vision methods to extract the visual features to represent the images. Then a classification step is performed by using machine learning algorithms. The research of image classi-ficationcan promote the development of web image retrieval, intelligent video sur-veillance, biometrics recognition and so on.In spite of so many applications, the current research on image classification can-not satisfy the demand for real applications. The semantic gap occurs between the low-level visual representation and the high-level human sense is the main prob-lem.The rationale behind this problem is the discriminative ability of the image classi-fication system.Thus, this dissertation focuses on the research of the discrimination enhanced image classification. The main research work and contributions are as fol-lows:1) The state-of-the-art low-level feature coding methods have been studied and low-level feature coding is performed by the relationship between two impor-tant factors which are low-level feature and codebook. Considering the cha-racteristics of these two factors, this dissertation proposes to use explicit addi-tive kernel maps to transform the low-level feature and codebook into a high-er feature space for separabability enhancement. The vector difference be-tween the low-level feature and codebook in the transformed feature space is utilized to describe their relationship.This dissertation proposes a generalized low-level feature coding method.The research illustrates that the proposed method has more powerful discriminative ability.The performance of image classification on the public dataset has been improved.2) The limitations of the existing image representation methods have been pointed out. In order to minimize the impact of the variability in image classi-fication, this dissertation proposes to model the latent factors that have the in-fluence onthe final classification performance by the analysis of the variabili- ty of image classification. Meanwhile, this dissertation proposes a discrimina-tive frameworkto enhance the discrimination of the image representation. The framework can represent each image by a low-dimensional feature vector based on partial least square method. This dramatically reduces the burdens of the classifier training and the feature storage. Because the framework uses the image labels, the final image representation has strong discriminative ability among different image classes. The effectiveness of the proposed method has been proved on the popular image classification datasets.3) An online discriminative parametric image similarity metric learning algo-rithm has been proposed. Based on the basic image representation framework, the proposed algorithm uses the pairwise constraints to learn the parametric image similarity metric.The pairwise constraintsencode theimage label prior so thatan image may have larger similarity between the images in the same class than those in the different classes after learning. Therefore, the discri-minative ability of the learned similarity measure is enhanced.Meanwhile, this dissertation proposes an online learning algorithm to deal with the large scale learning problem which is the consequence of the pairwise constraints. As shown in the experimental results, the proposed online learning algorithm achieves the promising performance and improves the efficiency of the rela-tive offline learning algorithm.4) A global and local training framework whichcan be applied to image classifi-cation has been proposed.Considering the moving object classification, this dissertation analyzes the distribution of the input space and points out that there are global and local information that can be used for classifier training. Using initial clustering and clusters refinement, this dissertation dividesthe input space into several local clusters. A trainedglobal classifier captures the global information while the localclassifiers can capture the local information. This frameworkcan deal with the problem of complex data distribution ofin-put feature space, especially for video surveillance. Theexperimental results and the real application system illustrate the advantages and the practicability ofthe proposed method respectively.