ILDA Based Writer Adaptation For Handwritten Chinese Character Recognition

Great progress has been achieved in the field of online hand- written Chinese character recognition (OHCCR) during the past 40 years, but recent researches on unconstrained cursive online handwriting recognition show that this problem is far from having been completely solved and the accuracy of identification is still not meet the user's requirement. The writer adaption is the process of converting a writer-independent system learned from the writer-independent dataset to a writer-dependent system, which is turned for a particular writer using a specific incremental data. As the writer adaptation is an online incremental learning process to learn the particular writing behavior and adaptively updating the classification model. This adaption has the potential advantage of significantly increasing recognition accuracies for a particular writer, which is very useful for a real world application. In the past, a number of writer adaptation handwriting recognition methods have been proposed. Unfortunately, all these methods are designed for small scale handwriting recognition problem, where the class number is relatively small; thus many of the adaptation methods are not practically applicable for handling large datasets with many classes, such as Chinese handwriting recognition problem involving thousands of classes and hundreds of thousands of training/testing handwritten samples.Although a number of researches on writer adaptation or ILDA were conducted, the ILDA based writer adaptation handwriting recognition remains unexploited. Motivated by this problem, we investigate how to adapt a writer independent recognizer to make it writer dependent based on the incremental learning of LDA model under the LDA based OLHCCR classification framework for the first time in this paper and apply it to the Cloud Computing input method invented by Jingroup Lab.We first provide a general incremental learning solution for LDA.The former incremental LDA for classification is too complex, since the sequential incremental learning condition and the chunk incremental learning condition are considered separately. In addition, for each case, the solution is divided into two cases depending on whether the new class sample is added or not. If the new class was introduced, the number of the new class must be 1. In contrast, our approach can solve all of the above situations using a uniform framework without restricting the number of newly introduced classes.Based on the general incremental learning solution for LDA, We then propose a weighted incremental linear discriminant analysis (WILDA) approach for writer adaptive handwriting recognition by taking into account the issue of uncertain number of incremental data for writer adaption in an online handwriting recognition application. And experimental results show that WILDA outperforms ILDA. Compared to ILDA, WILDA is more useful for a real world application.Experimental results show that both ILDA and WILDA are very effective to improve the recognition accuracy for particular writers, and WILDA outperforms ILDA. The experimental results indicate that the writer adaption using the WILDA approach can not only significantly increase the recognition accuracy for the particular writers but also have limited impact on the accuracy for the general writers.Finally, we design a writer adaptive handwriting recognition system which consists of the training phase to train a general baseline classifier, the writer adaptation phase using ILDA/WILDA, and the classification phase. And apply it to the Cloud handwriting input method invented by Jingroup Lab.