Classification and feature extraction methods with applications to image database retrieval.

摘要

For an image database, image Retrieval is to find the collection of images that interest its users. It is a difficult question in statistical analysis since image databases are often very large and of high dimensionality. In this dissertation, we will investigate how images can be classified and retrieved automatically, quickly and accurately.; In Chapter 1, we will give an overview of the problem. First we discuss some representations of color, texture and shape features. Secondly, we introduce the problem we aim to tackle. We consider classification as the fundamental problem while paying less attention to the sophisticated engineering details.; Chapter 2 talks about statistical models that have been proposed for classification problems. We discuss their advantages and disadvantages with respect to their applicability to image retrieval.; Chapter 3 proposes a new ensemble scheme Mutual-exclusive Boosting (MxBoost). Traditional ensemble schemes such as Bagging and Boosting can improve most classification methods, but they are computationally expensive, require large amount of storage and suffer from long classification time. MxBoost can improve classifier ensembles without increasing the ensemble size, therefore make it attractive for online classification problems where classification has to be performed quickly.; Chapter 4 discusses the issue of feature extraction. Image database is often of very high dimension. However, a user may be interested in only a few of them when s/he queries the database. Selecting a number of variables wisely can improve both classification accuracy as well as retrieval speed.; The last chapter proposes a new directional feature extraction method Spin Discriminant Analysis (SDA). Its connections with Linear Discriminant Analysis (LDA) and Parzen's Window method are established. Extensive experiments are carried out to show that SDA is an effective classification method.