Digital microscopy is an emerging technology which presents some challenges to a system designer. Scanning an entire slide at maximum resolution can produce an image gigabytes in size. Common scanning techniques produce visible seams in the resulting images. I have developed a system to acquire higher quality scans of slides at a higher magnification than currently available. I have also developed a user friendly method of navigating and annotating the images. This system is called The Urbana Virtual Microscope, and is found at http://histo.org.; In this thesis I have implemented a content based image retrieval system which allows searching for small regions within the large slides. I have compared three sets of image features: color features from a histogram of hue and saturation, structure features from a waterfilling algorithm on an edgemap, and texture features from wavelet analysis. I show that these three feature sets are independent and complement each other well, allowing a user to search for a wide variety of tissue categories.; The system also uses relevance feedback to give extra emphasis to the features which have high correlation among those selected by the user. I show that greater amounts of feedback produce more precise results.; Finally I developed a variation of hierarchical clustering which I call connected clustering. It groups neighboring images that are similar, allowing the system to treat them as a single unit. I show that this improves the response time of the system and reduces redundancy in the results displayed to the end-user. |