| Machine learning in computer vision is an emerging field aimed at improving the performance of computer vision systems. Learning-based systems are expected to provide a higher level of competence and greater generality. Recently model-based approaches have been used with machine learning techniques to enable automatic model acquisition from visual training data. On the other hand, the goal of improving the performance of computer vision systems has brought new challenges to the field of machine learning, for example, learning from structured descriptions, incremental learning, and learning with many classes.; This thesis presents a new approach that uses unsupervised learning, called on-line shape learning, to find a set of templates specific to the objects being outlined by the user. On-line learning gives the system improved performance with continued use by adjusting the clusters, and by creating a new cluster whenever an unusual shape is presented.; On-line shape learning is first applied to a unified model-based approach for detection, segmentation, and classification of objects in two-dimensional images. The learned templates allow intelligent search of templates for detection, the realistic initialization of object boundaries for segmentation, and the recognition of particular classes for classification. Keeping the model-based approach, a new neural network is created based on on-line shape learning. The proposed neural networks use unsupervised learning to cluster shapes for detection and segmentation in a hidden layer, and supervised learning to classify objects. These techniques are applied to the problem of classifying cell nuclei in cytological breast cancer images.; The second application is an object-based image retrieval system incorporating on-line shape learning. By utilizing user feedback, on-line shape learning can learn users' preference to give the user more control on the search criteria, and to allow the database system to learn the users' preferences. When the user looks for an object similar to one found previously, the system can give results using preference learned by feedback, instead of starting an initial search. One contribution of this thesis is that the proposed system is the first application to apply learning based on user feedback to indexing in information retrieval. |
