A new linguistic decoding method for online handwritten Chinese character recognition

In various application areas of pattern recognition, combining multiple classifiers is regarded as a method for achieving a substantial gain in performance of systems. The paper presents a method for handwritten Chinese character recognition to combine multiple classifiers based on statistics. Fusion strategies are discussed for providing a basis for combining classifiers. These combination strategies are experimentally tested on an online handwritten Chinese character recognition system. In our experiments, other combination approaches are also involved for comparison.

Handwritten Chinese character recognition (HCCR) is an important research field of pattern recognition, which has attracted extensive studies during the past decades. Recently convolutional neural network (CNN) based methods have achieved the state-of-the-art performance for handwritten Chinese character recognition. Nevertheless, handwritten Chinese character recognition is still limited to be effectively used in the actual environment due to the large-scale vocabulary and great diversity of handwriting style. In this paper, we constructed a handwritten Chinese character recognition service based on convolutional neural network, which tries to make effective use of handwritten based printed fonts and existing handwritten database. At the same time, the service can effectively collect more handwritten data to expand the training dataset, which makes it easy to adapt to the new handwriting styles. Meanwhile, We propose a multi-level recognition theory applied to online handwritten Chinese character recognition, which may improve the accuracy of handwritten Chinese character recognition and break the limitations of handwritten Chinese character recognition by identifying the structure of Chinese characters and possible stroke orders firstly. Furthermore, we try to apply the method of online character recognition to the offline character recognition based on the basic writing rules.

This paper presents a new approach to large-vocabulary online handwritten Chinese character recognition based on semi-tied covariance (STC) modeling. Detailed procedures are described for estimating the STC model parameters under both maximum likelihood (ML) and minimum classification error (MCE) criteria. Compared with the state-of-the-art modified quadratic discriminant function (MQDF) based classifiers, STC-based classifiers can achieve a better memory-accuracy trade-off, thus provide more flexibility in designing compact online handwritten Chinese character recognizers. Its usefulness has been confirmed and demonstrated by comparative experiments on popular Nakayosi and Kuchibue Japanese character databases.

The hierarchical nature of Chinese characters has inspired radical-based recognition, but radical segmentation from characters remains a challenge. We previously proposed a radical-based approach for on-line handwritten Chinese character recognition, which incorporates character structure knowledge into integrated radical segmentation and recognition, and performs well on characters of left-right and up-down structures (non-special structures). In this paper, we propose a statistical-classification-based method for detecting special radicals from special-structure characters. We design 19 binary classifiers for classifying candidate radicals (groups of strokes) hypothesized from the input character. Characters with special radicals detected are recognized using special-structure models, while those without special radicals are recognized using the models for non-special structures. We applied the recognition framework to 6,763 character classes, and achieved promising recognition performance in experiments.

This paper proposes an on-line handwritten Chinese character recognition system based on stroke-sequence feature extraction. The character to be recognized can be stroke-order and stroke-number free, tolerance for combined strokes, size flexible, but within the constraint of normal hand-writing. Firstly, the recognizer, using the finite state matching mechanism, is used to extract primitive strokes, represented as stroke string, from the input character. Secondly, the recognizer, using a modified dynamic programming matching method, is employed to perform recognition processes with the stroke-string features. Reference patterns have been generated 2500 Chinese characters with stroke-numbers ranging from 1 to 29. The recognition results are based upon the 1800 handwritten characters by 10 people. The obtained recognition rate is 94.5%, and the cumulative classification rate of choosing fourth most similar characters is up to 98.7%. In the last part, a secondary recognition mechanism is used to further tell apart the candidates involved. The final recognition rate may be promoted up to 99%.

Bipartite weighted matching for on-line handwritten Chinese character recognition

This paper presents a novel approach to writer adaptation using bottleneck features and discriminative linear regression for the recognition of online handwritten Chinese characters. First, bottleneck features extracted from a bottleneck layer of a deep neural network representing a nonlinear and discriminative transformation of the input features are verified to be much more effective in adaptation of writing styles than the conventional features after linear discriminant analysis transformation. Second, discriminative linear regression via a so-called sample separation margin based minimum classification error criterion is adopted for writer adaptation. The experiments on an in-house developed online Chinese handwriting corpus with a vocabulary of 15,167 characters and testing data collected from user inputs of Smartphones show that our proposed approach can achieve very significant improvements of recognition accuracy compared with a state-of-the-art adaptation approach for writer adaptation.

A new perspective: Recognizing online handwritten Chinese characters via 1-dimensional CNN

In handwritten Chinese character recognition, the performance of a system is largely dependent on the character normalization method. In this paper, a visual word density-based nonlinear normalization method is proposed for handwritten Chinese character recognition. The underlying rationality is that the density for each image pixel should be determined by the visual word around this pixel. Visual vocabulary is used for mapping from a visual word to a density value. The mapping vocabulary is learned to maximize the ratio of the between-class variation and the within-class variation. Feature extraction is involved in the optimization stage, hence the proposed normalization method is beneficial for the following feature extraction. Furthermore, the proposed method can be applied to some other image classification problems in which scene character recognition is tried in this paper. Experimental results on one constrained handwriting database (CASIA) and one unconstrained handwriting database (CASIA-HWDB1.1) demonstrate that the proposed method outperforms the start-of-the-art methods. Experiments on scene character databases chars74k and ICDAR03-CH show that the proposed method is promising for some image classification problems.

Model-based stroke extraction and matching for handwritten Chinese character recognition

Automating handwritten character recognition (HCR) for Sinhalese is still an emerging field, as Sri Lanka is the only country in the world that uses Sinhala as the national language. The Sinhala alphabet includes 60 characters that appear to be fairly complex when compared to many other languages. Although approximately 25 research studies have been conducted since 1990s to recognize handwritten Sinhala characters, there is still a large room for improvement. We have investigated two different methods for handwritten Sinhala character recognition: (i) total pixel length-based method which was newly derived by assuming the differences among characters in pixel length and (ii) Convolution Neural Networks (CNN) based classification, to overcome the limitations in the primitive methods. Multiple preprocessing techniques were used in the pixel length-based method while only a resizing technique was used in the CNN based method. The total pixel length and CNN based methods were tested using a dataset of 55 characters, each with 33 samples. The result of the pixel length method is imperfect as distinct characters may hold the same length. Therefore, we abode this method for further processing in future with more improvements. But the proposed CNN classification produces an accuracy of 85.71% and when compared to current studies, this can be considered as the best result achieved for 55 characters among total 60 characters, with nearly 10% of accuracy improvement.

Multi-stroke relaxation matching method for handwritten chinese character recognition

Chinese characters are constructed by basic strokes based on structural rules. In handwritten characters, the shapes of the strokes may vary to some extent, but the spatial relations and geometric configurations of the strokes are usually maintained. Therefore these spatial relations and configurations could be regarded as invariant features and could be used in the recognition of handwritten Chinese characters. In this paper, we investigate the structural knowledge in Chinese characters and propose the stroke spatial relationship representation (SSRR) to describe Chinese characters. An On-Line Chinese Character Recognition (OLCCR) method using the SSRR is also presented. With SSRR, each character is processed and is represented by an attribute graph. The process of character recognition is thereby transformed into a graph matching problem. After careful analysis, the basic spatial relationship between strokes can be characterized into five classes. A bitwise representation is adopted in the design of the data structure to reduce storage requirements and to speed up character matching. The strategy of hierarchical search in the preclassification improves the recognition speed. Basically, the attribute graph model is a generalized character representation that provides a useful and convenient representation for newly added characters in an OLCCR system with automatic learning capability. The significance of the structural approach of character recognition using spatial relationships is analyzed and is proved by experiments. Realistic testing is provided to show the effectiveness of the proposed method.

An efficient method of Chinese handwritten address character string segmentation and recognition is presented. First, an address string image is presegmented into several radicals using stroke extraction and stroke mergence. Next, the radical series obtained by presegmentation merge into different character image series according to different merging paths. After that, the optimal merging path is selected using recognition and semantic information. The recognition information is given by the character classifier. The semantic information is obtained from large scale address database containing more than one hundred thousand address items. Finally, the optimal recognition results of the character image series which are combined by radical series according to the optimal merging paths are obtained. In experiments on 897 mail images, the proposed method achieves correct rate of 85 percent while the error rate is 15 percent.

Highly structured neural networks like the Time-Delay Neural Network (TDNN) can achieve very high recognition accuracies in real world applications like on-line handwritten character and speech recognition systems. Achieving the best possible performance greatly depends on the optimization of all structural parameters for the given task and amount of training data. We propose an Automatic Structure Optimization (ASO) algorithm that avoids time-consuming manual optimization and apply it to Multi State Time-Delay Neural Networks (MSTDNNs), a recent extension of the TDNN. We show that MSTDNNs are a very powerful approach to on-line handwritten character and word recognition and that the ASO algorithm can automatically structure this type of architecture efficiently in a single training run.KeywordsSpeech RecognitionState UnitHide UnitDynamic Time WarpingOutput UnitThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.