Journal of Pattern Recognition Research

Automatic recognition of broken Thai characters represents one of the biggest challenges in some applications such as computerized restoration of Thai text documents. We propose a novel solution based on set-partitioning of the broken character pieces. The basic idea is to find the optimal set-partition of broken pieces that maximizes the likelihood that each block of the partition is well-grouped and resembles a character in terms of pattern and sizing. Our approach is tolerant to presence of noise pieces in the image. In addition, we devise a mechanism based on A* search to align a sequence of Thai characters into Thai words using a dictionary modeled as N-grams graph structure. Experiments based on this framework have been performed and the results are very promising.