Handwriting is a natural way to communicate and record information. Despite more than four decades of intensive research, offline unconstrained cursive handwriting recognition is still an unresolved problem. Touched cursive characters are not uncommon and are the main cause of low segmentation and recognition accuracy.
This thesis presents enhanced approaches for image pre-processing, touched character segmentation and feature extraction for character recognition. Enhanced pre-processing techniques include noise detection and removal, image skew estimation and detection of handwritten and machine-printed text.
Noise detection is based on a connected component analysis scheme while geometrical features are employed to estimate image skew angle. Local and global features are analysed and employed to distinguish between handwritten and machine-printed text. However, touched character segmentation is the main focus of this thesis. In this regard, two enhanced touched character segmentation techniques based on genetic algorithms and pixel intensities are proposed and evaluated.
For character recognition, statistical and structural features are extracted and a fused technique is proposed. All techniques are evaluated on real-world benchmark IAM data that facilitated comparison in the state of the art. Favourable accuracy for each phase is achieved and reported in this research with high speed and minimum computational complexity.
4 July 2011