With the rapid advancement of the computing and Internet technology, theamount of various multimedia data including digital images increases in a surprisingspeed. Due to the nature of unstructured arrangement for those data, it becomes verydifficult for people to search for relevant information over the Internet. While peoplebenefit a lot from the advent of information digitized technology, they also have toface the dreadful dream about how to analyze, store and retrieve the huge amount ofdata efficiently and effectively, especially for those multimedia data. Thisdissertation is dedicated to the study on the key techniques of content-based imageanalysis and retrieval.We first give an introduction to the state of the art of content-based imageanalysis and retrieval technology. In the introduction, we review the analysis andrepresentation of image content, some common techniques for content based imageretrieval, and other related issues. We conclude that image content representationand image feature match should be two key research topics in the dissertation.As for image content representation, we first discuss the extraction of textureand color features for an image. The extraction of texture feature is based on SARtexture analysis model. We combine MRSAR with non-symmetric and non-causalneighboring set to estimate the unknown parameters of the model. The extraction ofcolor feature derives from a dominant color index approach we proposed, which isbased on a multi-resolution fuzzy partition policy for an image. After thecombination of our method with global color histogram, we implement an effectivecolor retrieval approach, which makes full use of not only the color distribution, butalso the color spatial distribution.Furthermore, we propose a novel approach to the unsupervised extraction ofdiscriminant regions in an image. The discriminant region is obtained after theapplication of SOM learning and SOM reduction. In the dissertation, we present anovel two-phase reduction algorithm for a SOM, and a cluster-validity analysisindex to determine the optimum number of regions in an image. Based on theanalysis of the computation complexity of our algorithm, some optimizationtechniques are considered which can obviously reduce the computation time of thealgorithm. We also present a vote algorithm that groups pixels in the image intorelevant discriminant regions. The obtained discriminant regions for an image arealways regions of interests in the image, which can be applied to not only imageretrieval, but also other applications such as image classification and filter, etc.As for image feature match, we mainly study the fast computation of thedistance measurement between two features, and relevance feedback technology in...
|