Background Features for HMMs-based Off-line Handwritten Character Recognition

Feature extraction is one of the most important factors for recognition system based on hidden Markov models (HMMs). This paper presents a new approach by using background feature and HMMs for off-line handwritten character recognition. The background feature for white pixels (also called background pixels) is based on the Freeman code with eight directions, which is an improvement of Alceu´s method with four directions. Experimental results for off-line handwritten Chinese legal amount show the validity of the new approach. The recognition rate is about 1.8~4.9% higher than that of Alceu´s method with the same HMMs topology. For the new approach within the tested topologies, the highest recognition rate can be 96.39%.