Performance of curvelets, dual-tree complex wavelet and discrete wavelet transform in handwritten word classification

Handwritten words classification is a difficult task due to the high variability and uncertainty of human writing styles. The aim of this work is to Performance of curvelets, dual-tree complex wavelet and discrete wavelet transform in handwritten words classification. Curvelet transform, Dual-Tree complex wavelet transform (DTCWT), Haar, Daubechies, Coiflets, Symlet, Discrete Meyr, Biothogonal and reverse Biothogonal are used in this investigation. Three to four wavelets are chosen randomly from each wavelet family. A dataset of 534 handwritten Latin (HL) and Arabic (HA) word images out of 1068 (267 of each script) are used for training and the remaining is for testing. Energy and entropy are computed at different decomposition sub-bands. Support vector machines (SVM) and 1-NN are used for classification. An exhaustive experimentation is carried out on word images downloaded from IAM Handwriting Database for Latin words and from IFN/ENIT-database for Arabic handwritten words. The experiments showed that the best identification results are achieved using Curvelet transform followed by DTCWT.