Laplacian and Bilaplacian Based Features for Shape Classification

Finite difference schemes of various eigenvalue problems are used to generate size, rotation, and translation invariant sets of features for shape recognition and classification of binary images. These feature sets are based on the eigenvalues of the Dirichlet Laplacian, the clamped plate problem, and the buckling problem. The stability and effectiveness of these features is demonstrated by using them in the classification of 6 types of computer generated and hand-drawn shapes. The classification was done using 4 to 20 features fed to simple feed-forward neural networks trained using the backpropagation algorithm. All features performed very well and correct classification rates of up to 99.7% were achieved on the computer generated shapes and 97.2% on the hand-drawn shapes.