SH-MQDF based calligraphic character recognition

Chinese Calligraphy draws a lot of attention for its beauty and elegance. Until now, lot of people don't know the semantic meaning of many calligraphic characters because they were written in ancient times. Technologies which can help users to recognize the unknown calligraphic characters are urgently required. However recognizing calligraphic characters is a challenging work due to the reasons: (1)complication. There are at least five different styles of the same character; (2)Deformation. Sometimes deliberate deformation in writing is performed, which makes the recognition even harder. In this paper, we present an approach for the recognition of calligraphic characters. Modified quadratic discriminant function (MQDF), which was the state-of-the-art classifier used in handwriting recognition, is introduced in recognizing calligraphic characters. Due to the high expenses of storage and computation of MQDF on large scale data, we use spectral hashing(SH) to dynamically reduce the whole training set to a much more relevant but much smaller set, called the approximate set. MQDF is then trained on the approximate set dynamically and classification is performed. We evaluated our approach on the calligraphic character dictionary (CCD), which is an expert labeled calligraphic character database. Our experiments show that our approach reduces the training time of classifier but doesn't lower the accuracies of recognition much. Furthermore, our approach outperforms the existing calligraphic character recognition methods on CCD.