We introduce a method for semi-automatic classification of 2D shapes based on their Curvature Scale Space (CSS) image representation. Every object is represented by the maxima of the curvature zero-crossing contours of its CSS image. The similarity between two different shapes is then expressed by a real value which is the result of comparing their CSS image representations.
Having a training set of classified shapes, the similarity between an unclassified input and a subset of classified shapes is measured. The k-NNR method is then used to identify the classes most similar to the class of the input shape.
We have used this method in a real world application to find, for an unknown leaf, similar classes from a database of classified leaf images representing different varieties of chrysanthemum. The task is to find out whether the unknown leaf belongs to one of the existing varieties or it represents a new variety. The system finds the most similar varieties to the input and allows the user to make the final decision.
We have tested our method on a prototype database of 400 leaf images from 40 different varieties. The results have indicated a promising performance of the system.