Serialization And Recognition Method Of Online Handwritten Character

With the improvement of online handwritten character recognition methods, their complexity has also been greatly increased. Nevertheless, even for the state-of-the-art algorithm, i.e., the Convolutional Neural Network(CNN) model, there are some users for which only about 60% of the recognition rate have been achieved. Since the complicate of existing model, it is really hard to improve the performance of the recognition algorithms for these users through online and real time training. To address this issue, it is necessary to seek a novel method that can be learnt incrementally with real time and can adapt to the user automatically, while its complexity must be limited for running on intelligent mobile device, which is a new challenge for online handwritten character recognition and is still not well researched.To address above issue, this thesis proposes a handwritten character recognition method bases on segmenting handwritten characters into basic and language independent natural stroke sequences, which are then further converted into ASCII character strings by using pre-defined coding book and natural stroke recognition method. The proposed method consists of three parts: At first, to get robust sequences of natural strokes, all online-collected characters are preprocessed and are segmented by predesigned rules and vector walking method. Then, the segmented results are converted into ASCII strings via pre-given coding book. The string matching method based on minimum edit distance computing is used as our handwritten character recognition method. To further improve the recognition rate, the fuzzy scale measure of strokes is extracted with a 3-dimensional vector for each stroke. The candidate results extracted by minimum edit distance method are finally re-ranked and output as the result of our fuzzy-scale based recognition method.To verify the effectiveness of proposed method, this thesis uses the standard corpus HIT-OR3 C and separates it into training and testing parts. Experiments are conducted for evaluating the pre-processing, segmentation and matching based recognition methods. The results show that, using the preprocessing and the vector walking segmentation method with 24 types of natural strokes, the minimum edit distance based method obtain 55.85% and 89.58% recognition rates for top 1 and top 10 results respectively. Use two writing person dependent samples as the training set, and use another character set writing by the same person as the testing set, the recognit ion rates reach 65.92% and 91.90% for top 1 and top 10 candidates. By using fuzzy-scale based re-ranking method on the first 20 candidates of each testing character, the recognition performances are greatly improved, 79.87% and 94.26% for top 1 and top 10 candidates are reached. These results prove that, compared with the traditional methods, our method can effectively reduce the training time, add new samples into existing models in real time, while achieve acceptable recognition results, which shows that this method have great advantage in incremental learning and user adapting tasks.