Parsimonious Modeling Methods For Large-Vocabulary Offline Chinese Handwriting Recognition

Offline Chinese handwriting recognition is a challenge topic due to large vocabu-lary and unrestrained writing styles,two tasks are explicitly defined,including offline handwritten Chinese character recognition and offline handwritten Chinese text recog-nition.Recently,great success has been achieved in offline Chinese handwriting recog-nition by using deep learning methods.However,due to the large-vocabulary Chinese characters(Overall simplified Chines characters more than 27000,overall complex tra-ditional Chinese characters more than 100000),many problem have arisen.First,a high demand of memory and computation is required.Second,a lot of data needs to be used for both common or uncomon characters.Third,system unable to recognize Out-of-Vocabulary(OOV)characters and newly created characters.To address the problem of high demand of memory and computaion,we present parsimonious HMM(PHMM)via Two-steps algorithms which can fully utilize the simi-larities among different Chinese characters.On offline handwritten Chinese text recog-nition task,compared with traditional HMM system,PHMM not only lead to a compact model but also improve the recognition accuracy and reduce the decoding time.To address the problem that system needs a lot of data and is unable to recog-nize OOV characters,we propose a novel radical analysis network with densely con-nected architecture(DenseRAN)to analyze Chinese character radicals and its two-dimensional structures simultaneously.On offline handwritten Chinese character and text recognition tasks,the manner of treating a Chinese character as a composition of two-dimensional structures and radicals can reduce the size of vocabulary and enable DenseRAN possess the capability of recognizing unseen Chinese character classes,only if the corresponding radicals have been seen in training set.