Offline shape recognition using flexible DCT grid

Extracting features out of binary shapes which are robust against scaling, noise, and rigid & elastic deformations is the purpose of this paper. The authors have utilized the amazing energy compaction property of Discrete Cosine Transform, applied to variable size, non-overlapping blocks of input image, containing the shape to capture features (hereby called the signature of the shape). By keeping the number of these non-overlapping blocks the same, irrespective of the size of input image, is good against scaling, while applying various morphological operations on black and white versions of the shape and then picking a few lower frequency components makes the scheme work against noise and deformations. Extracted signatures show amazing inter-class variation and intra-class similarity when classified using minimum distance and k-NN classifiers. The proposed system has shown great accuracy on GREC 2003 shape dataset (http://www.cvc.uab.es) [23], MPEG7 silhouette shape database (http://www.imageprocessingplace.com/root_files_V3/image_databases.htm) [25] and TOSCA tools shape dataset (http://tosca.cs.technion.ac.il/book/resources_data.html) [26].