Entropy based shape discrimination

Object recognition based on morphological properties is a routine task that a human being performs with intelligence. In an attempt to integrate such capability in an automatic vision system, this work presents a new approach to shape understanding. This approach is based on using the shape characteristic function, which reduces the task of 2-D shape analysis to 1-D function analysis. Moreover, shape normalization with respect to location, scale, and orientation becomes a reasonable task. Based on the normalized shape, entropy, being the self information in a probability density, is shown to measure the degree of circularity in any shape. Other shape properties are evaluated using the mutual information concept which measures the degree of similarity between different shapes. Shape recognition and classification is achieved once prototypes are selected. In the process of shape discrimination, the adjustment of the size of samples is achieved using a nearest neighbor substitution algorithm in order to maximize the information content in a shape. An introduction for analyzing 3-D objects based on their 2-D cross-sections is given for future work. Finally, different types of applications are considered for testing the descriptor's performance in analyzing global as well as local variations.