A Statistical Framework for Scale, Rotation, and Translation Invariant Object Recognition

The process of feature extraction is fundamental to object recognition process, The paper proposes techniques for computing boundary and region descriptors, which are efficient and work well even in the presence of noise. For boundary descriptors, shape is extracted through morphological algorithms and boundary of the object is traced, after they have been suitably binarized and noise has been removed. Boundary obtained in this way is manipulated to obtain boundary descriptors like chain code, shape numbers, Fourier descriptors applied on signatures (both distance from centroid and coordinates), moments and other scalar descriptor. Similarly region descriptors like moments, skeleton and scalar descriptors are calculated. The statistical patterns obtained from these descriptors are matched against already calculated patterns stored in database for recognizing object, where pattern is a set of descriptors. The patterns proposed in the paper are invariant to scaling, rotation and translation transformation.