Research On Some Problems In Handwritten Chinese Text Recognition

Handwritten Chinese text recognition is to use a computer to automatically recognize an input of handwritten text and extract its corresponding text content.With the arrival of the fourth industrial revolution with artificial intelligence as its core driving force,handwriting recognition now plays an increasingly important role in many human-computer interactions:such as automatic express delivery,intelligent examinations,information entry,and so on.Considering China's large population and influential position in the world.It is important to study the recognition of Chinese text.However,compared with isolated character recognition,continuous character recognition is a kind of sequence solving problem,and how to effectively model this is a difficult point of research.Since the 20th century,researchers have proposed many solutions based on statistical modeling methods.These methods can be divided into over-segmentation methods and segmentation-free methods according to whether pre-segmentation is required.On the one hand,accurately finding the segmentation point often becomes the performance bottleneck of the over-segmentation methods,on the other hand,limited by the ability of the traditional classifier,the recognition performance of different methods has not made much progress.In recent years,with the availability of big data and the rapid advancement of deep learning technology,the performance of continuous handwriting recognition has been greatly improved,whether it is the over-segmentation methods or the segmentation-free methods.Although the existing methods based on deep learning have obtained significant performance improvements compared to traditional methods,there are still some areas worthy of further study.The first is how to combine traditional sequence models and deep learning techniques more effectively to realize the full use of training data,especially a large number of isolated character data.The second is that there are many kinds of Chinese characters,some of which are highly similar.How to leverage mathematical models to automatically detect and cluster similar basic units between these characters to improve the modeling ability of the neural network and reduce the ambiguity of the decoder in recognition.The third is to consider the the different writing styles in an unconstrained environment,how to use adaptive technology to realize the adaptation of writers,and further improve recognition performance.Finally,how to design a more compact network model to achieve low storage consumption and low computational classifiers under high category numbers.This paper studies new solutions to the previously mentioned challenges in handwritten Chinese text recognition.It is worth mentioning that the new methods studied for certain problems have a certain universality and can be naturally extended to other fields.First,it addresses the shortcomings of the limited capacity of classifiers and insufficient utilization of training data in the previous mainstream sequence models.This paper proposes to use a combination of neural network and hidden Markov model to deal with Chinese continuous handwritten text recognition.This paper discusses four aspects of the effectiveness of the proposed method in detail,namely,feature extraction,classifier design,training criteria,and language model selection.The modeling method based on the hidden Markov model can make full use of the isolated character samples of the training set,improve the data utilization rate,and relieve the deficiencies caused by the lack of continuous text lines in the training set.At the same time,combining the neural network classification model and language model,we achieve a recognition rate that is higher than those of other mainstream methods.Secondly,in order to solve the problem of many Chinese character categories and high similarity,based on the built CNN-HMM system,this paper introduces a state tying technology to cluster similar parts in Chinese characters,thereby reducing the total number of states and improving the training and discriminative ability of the convolutional neural network.Furthermore,it reduces the decoding ambiguity caused by the inconsistent scores of similar parts in the decoding process.To tackle the difficulty of changing writing styles,we introduce adaptive layers in the convolutional neural network.This combines the state tying and unsupervised adaptive technology to reduce the performance loss caused by the writing style.Moreover,in order to reduce the delay problem caused by the aforementioned adaptive technology,this paper proposes a novel fast adaptive technology based on an auxiliary writing style extraction network.The extracted writing information is imported into the baseline recognition network to achieve the purpose of adaptation.Finally,in view of the shortcomings of storage consumption and high calculation amount in convolutional neural network of high class number,this paper proposes a distillation process combining structure and knowledge.The proposed method,through the organic combination of structural analysis,compact convolution development,and knowledge distillation,effectively constructed a more compact classification model on the premise of barely any loss of performance.This method is not only well applied in the task of continuous Chinese handwriting recognition considered in this paper,but also well verified in other mainstream classification models and tasks.