HMM Inception-Restnet Network Classifier for Manchu Word Recognition

The recognition of Manchu words is of great significance in studying and protecting Manchu culture and history. This paper proposes a Manchu word recognition system, using hidden Markov models (HMM) and Inception-Resnet Network hybrid architectures which shows superior performance on recognize tasks. The skeletonization and projection method was used to divide a Manchu word into character groups or characters. We select 280 characters or character groups from all shape of Manchu characters as the basic dataset, and then expand the Manchu character image dataset using several transformations. We adopt the soft pooling instead of the combination of max-pooling and avg-pooling in the revised Inception-Resnet network aiming at the small character images recognition, and propose a classifier for recognizing the segmentations of Manchu word. The experiments indicate that the average accuracy of this method is 0.9675 for the segmentations. Furthermore, the proposed method has no special requirements for the size of Manchu words.