Vector contour representation for object recognition in neural networks

The authors introduce a new method of describing object features for neural network based classification. The angle information between boundary pixels of the binary image of an object is extracted and preprocessed to form the vector contour representation (VCR). The VCRs of different objects are then used for pattern classification. The Euclidean distance between two VCRs is used as the similarity measure during classification. Scaling- and translation-invariant object recognition can be achieved. Partial or complete rotation invariance can also be obtained if desired. The usefulness of the vector contour representation for object recognition is demonstrated through the test results