Abstract
This thesis is concerned with the problem of shape similarity retrieval in image databases. Curvature Scale Space (CSS) image representation is examined for this purpose. It consists of several arch shape contours representing the inflection points of the shape as it is smoothed. The maxima of these contours are used to represent a shape. In order to make the representation more reliable and also reflection invariant, the conventional matching algorithm is modified. The method is then tested on a database of 1100 images of marine creatures, where the advantages and shortcomings of the method are discovered. One of the main advantages of the method is the existence of local support. This enables us to deal with the main shortcoming of the method which appears in case of shapes with shallow concavities. Several approaches are suggested and implemented to overcome the problem of shallow concavities. The shape is segmented using its CSS image, and more information is extracted from the segmented shape which is then used to enrich the representation. The matching algorithm is also modified to accommodate the new information. Each segment of the shape corresponds to a contour and consequently a maximum of the CSS image. In one approach the normalised average curvature on each segment is used together with the maxima of the CSS image to represent the shape. In another approach the segments are examined at different levels of scale and for each segment, the level of scale where it is converted to a straight line is determined. Using this information along with the maxima of the CSS image yields to the best results. Both inflection points and corners are considered as end-points in different approaches. There is less information about the global appearance of a shape in its CSS image. A small number of global parameters are included and used for indexing to obtain even better results. The problem of evaluation of similarity retrieval methods is addressed. In order to evaluate different approaches, a set of classified shapes are introduced and the performance measure of each method is calculated using this database. In another approach, a subjective evaluation of the method is presented based on the judgements made by human subjects. The method is also tested on a real-world application where the task is to help the users find out whether an unknown leaf belongs to one of the existing varieties or whether it represents a new variety. A web-demo of the work is also prepared and is available in the below mentioned address. The boundary contours of the marine animals images can be obtained from this page.