The development of multiscale techniques for object based image retrieval is the central aim of this dissertation. The area of Content Based Image Retrieval (CBIR) has received considerable attention in recent years owing to the large growth of digital image libraries. Individuals who seek to exploit such retrieval systems will frequently express their objectives in high-level semantics while the search algorithms require low-level feature description. Consequentially, additional techniques are required both prior to, and during the retrieval process to ensure the narrowing of this gap and hence facilitating successful retrieval.

This study introduces new algorithms for object segmentation and classification over multiple scales which allow for improved image content description. The majority of retrieval objectives can be accurately expressed through object content description. A new unsupervised image segmentation technique is developed which automatically determines object classes. During the process, models characterizing the colour and texture properties of each object class are obtained in the multiscale setting of a Hidden Markov Tree. This framework is based on the complex wavelet decomposition of a given image. The unsupervised Mean Shift Procedure is used to determine the number of object classes. Most unsupervised techniques are difficult to evaluate due to a lack of ground data on which evaluation of performance can be obtained. However, this thesis also presents a novel method for evaluating multiple segmentation techniques without ground truth through the use of a psychovisual testing procedure.

The object class models obtained during the segmentation phase form the basis of multiscale object feature descriptors for retrieval purposes. In addition, efforts have also focused on the development of techniques to aid successful retrieval through increased levels of computer-human interaction using relevance feedback algorithms. These techniques are designed to improve performance through relevance judgements on the accuracy of retrieved images. The use of Support Vector Machine Novelty Detection is proposed as a method of incorporating these relevance judgements into the retrieval ranking process.

Hence the techniques developed in this project provide a general application-orientated framework in which multiscale segmentation and classification techniques can be combined with relevance feedback algorithms to achieve effective and efficient retrieval.