Codec Network And GAN Based Handwritten Chinese Character Recognition And Generation

The development of writing was a major milestone in human civilization,allowing us to record and communicate information beyond the limits of human memory.With the rise of electronic devices,text recognition has become more and more important in human-computer interaction.Electronic devices can recognize printed text and even convert text on images into electronic formats,but such tasks are somewhat challenging due to the large variation between samples and the lack of standards.Handwritten Chinese characters can be divided into two categories according to whether they are written online or offline.Online handwritten Chinese character recognition technology has been very mature and widely used.Offline handwritten Chinese characters are stored in the form of images,which is more difficult to recognize,the recognition accuracy still needs to be improved,and the technology still needs to be more perfect.Along with the advancement of Chinese character recognition technology,the widespread use of electronic text has brought great convenience to people.The production of electronic text is inseparable from electronic fonts.However,traditional fonts often have a small number of fonts,and font design is a time-consuming and labor-intensive task.The generation of handwritten Chinese characters has broad prospects in the field of font design,and can greatly enrich the content of electronic fonts.The main work of this paper includes.(1)This paper proposes an offline handwritten Chinese character recognition network based on Ideographic Description Sequence(IDS).In recent years,the recognition of handwritten Chinese characters has become a popular research topic in the field of pattern recognition.The similarity between different categories of handwritten Chinese characters makes it limiting to distinguish them based on visual features alone,and simply increasing the depth or width of the network may not be effective.To address this challenge,the proposed model uses an encoder-decoder network to decompose handwritten characters into IDS sequences,combined with a convolutional classification network to predict the character’s classification.IDS represents the various parts that make up Chinese characters,and it can even represent Chinese characters that have not been seen before.At the same time,the convolutional classification network combines the center loss and cross-entropy loss during training to maintain the distance between categories while increasing the compactness within categories.Experimental results on CASIA-HWDB 1.0-1.1 and ICDAR 2013 datasets demonstrate the effectiveness of the above method.(2)This paper proposes a method for handwritten Chinese font generation based on sub-pixel generative adversarial networks.The emergence of generative adversarial networks provides a new idea for solving font generation tasks.However,generating Chinese fonts is more difficult than generating English fonts due to the complexity of Chinese characters in structure and details.In order to maintain the consistency of the generated font content,inspired by the task of handwritten Chinese character recognition,a convolutional classification network is introduced into the loss function of the generative adversarial network.Meanwhile,in order to improve the quality of Chinese font generation,some researchers try to use more complex network structures.These complex networks are difficult to train and the quality of the generated images is limited.The handwritten Chinese font generation method based on the sub-pixel generative confrontation network replaces the traditional transposed convolution with sub-pixel convolution,fully utilizes the information in the feature map,and reduces the interference of 0 pixels.At the same time,adding a residual network at the bottleneck of the generator can effectively utilize the underlying features and reduce the loss of feature information.