A Tibetan Component Representation Learning Method for Online Handwritten Tibetan Character Recognition

This paper presents a Tibetan component representation learning method for component-based online handwritten Tibetan character recognition. In conventional methods, we designed features manually for Tibetan components. The hand-crafted features are often incomplete and decrease the component recognition accuracy, which influences component-based character recognition performance. To overcome the deficiency, we use three layer deep belief networks to learn automatically representation features for components. Restricted Boltzmann machine is used to construct each hidden layer. The weight parameters of the networks are optimized by greedy layer-wise learning algorithm. Then we combine representation learning based component classifier into our previous integrated segmentation and recognition framework. Finally we add syllable association module to improve the handwriting input speed. Experimental results on MRG-OHTC database show that the component representation learning method gives the promising performance. The proposed method achieves the component-level and character-level recognition rates of 94.78% and 94.09%.