Research On Offline Handwritten Chinese Character Recognition Under The Guidance Of Print

Chinese characters have the characteristics of many categories and many similar char-acters,and the writing styles of Chinese characters are different,which leads to the ex-istence of exploration space for offline handwritten Chinese character recognition.In recent years,deep learning has developed rapidly,and the deep convolutional neural net-work model plays an important role in offline handwritten Chinese character recognition,but the model usually requires a large number of training samples.For handwritten Chi-nese characters,in addition to commonly used characters,the samples of other Chinese characters are not sufficient,which brings challenges to the recognition task.In response to the above problems,this paper uses the print to help and guide the recognition of offline handwritten Chinese characters from a more standardized point of view.The main work of this paper is as follows:1.Propose a retrieval network model based on the print guidance for offline hand-written Chinese character recognition.The model adds a retrieval branch on the basis of the classification network,and at the same time introduces the print dataset as gallery dataset.In the gallery dataset,the most similar print sample is retrieved for offline hand-written Chinese character,and the category of the print sample is used as the prediction result.Since the network achieves recognition work by comparing the similarity between samples,it has advantages in few-shot learning and zero-shot learning.2.Propose a model improvement method based on metric learning to make the fea-tures extracted by the model have a better spatial distribution.This method adds a metric learning branch to the retrieval network model,so that the distribution of samples of the same category is more concentrated,and the distribution of samples of different categories is more distant,thereby improving the robustness of the model.The final model achieves a recognition accuracy of 97.69% on the ICDAR-2013 offline handwritten Chinese char-acter competition dataset.Compared with the existing related research,this result has a certain improvement.3.Propose a three-stage experimental method based on the print guidance,and extend the idea of using print dataset to guide handwriting recognition to offline handwritten Chinese text recognition.In the first and third stages,the method uses a custom-generated offline handwritten Chinese text dataset to train convolutional recurrent neural network.In the second stage,the constructed single-character handwritten Chinese text dataset and the corresponding single-character print text dataset are constructed.The collection is sent to the network,and only the network convolution module is trained to make it have stronger feature extraction capabilities,thereby improving the recognition effect of the model.In the end,the experiment has proved the effectiveness of this method.