Baseline-independent feature extraction for Arabic writing

XIE Xudong,LI Ning,PENG Liangrui,DING Xiaoqing
2012-01-01
Abstract:An off-line handwritten Arabic recognition system without pre-segmentation was developed based on a hidden Markov model(HMM).Different pre-processing approaches are combined to extract a set of baseline-independent features.The original images are height normalized with thinning and contouring operations.24 features are extracted in a sliding window which is shifted along the word image from right to left.159 models are built up,with the number of states in each model depending on whether there is a ligature.Tests conducted on the benchmark IFN/ENIT database gave a recognition rate of 94.5%.The tests show that the features in this work make good use of the relationships between adjacent characters and are robust when the word image is shifted up or down,even with different handwriting widths.The features then emphasize the information in small strokes.
What problem does this paper attempt to address?