Research On Generation Technology Of Mongolian Handwritten Samples Based On Generative Adversarial Network

In recent years,with the rapid development of digital technology,a massive of Mongolian handwritten literature resources need to be digitally preserved,which makes the study of Mongolian offline handwritten recognition very important.According to word formation of Mongolian,its vocabulary can be attained to millions,and the commonly used vocabulary is also hundreds of thousands.Therefore,it is very hard to build a Mongolian offline handwritten word images datasets manually which contains all vocabularies.As a result,the research on Mongolian offline handwritten recognition faces certain difficulties.Thus,in order to solve this problem,it is necessary to study the automatic generation technology of Mongolian handwritten samples.In the field of handwritten text image generation,the methods based on Generative Adversarial Networks(GANs)are usually used.Among them,Pix2 Pix and Zi2 Zi methods need paired training data.For example,two images of the same word in different forms,such as a printed image and a handwritten image,these two images are called a pair of paired training data.However,it is difficult to obtain a large amount of paired training data in practice,which makes the training of related models more difficult.Aiming at the problems of diverse Mongolian writing styles,severe handwriting distortion,and difficulty in obtaining paired training data,this thesis proposes a method that can automatically generate Mongolian handwritten samples of any vocabulary,so as to help collect Mongolian offline handwritten word images datasets which contains all vocabularies.The main research work of this thesis is as follows:(1)Aiming at the problem of difficulty in obtaining a large number of paired training data,this thesis adopts the Cycle-Consistent Generative Adversarial Networks(CycleGAN)model to avoid the problem of obtaining paired training data,and realizes the automatic generation of Mongolian handwritten samples.The CycleGAN model aims to learn the mapping between styles of two data domain,rather than the mapping between styles of two specific data,and the model introduces the concept of cycle consistency to ensure that the learned mapping is a one-to-one mapping.Therefore,the CycleGAN model can automatically convert images to images without paired training data.Based on the CycleGAN model,this thesis uses unpaired training data to learn the writing style of Mongolian on a certain scale dataset of Mongolian offline handwritten word images,and transfer the obtained writing style to the word images of standard writing(Hawang).Then the word images of standard writing are converted into the handwritten word images with corresponding writing style.Thus,it can realize the automatic generation of Mongolian handwritten samples.The experimental results show that the CycleGAN model can generate Mongolian handwritten word images of any vocabulary with a certain writing style,which solves the problem of automatic generation of Mongolian handwritten samples to a certain extent.However,the writing style of the handwritten word images generated by the CycleGAN model is relatively single.(2)Aiming at the problem of the relatively single writing style of generated images in the CycleGAN model,this thesis proposes a method of automatic generation of Mongolian handwritten samples based on GANwriting.Specifically,a modification of original GANwriting has been presented by integrating a perceptual adversarial loss into the original GANwriting model in this thesis.For the convenience of description,we will introduce the GANwriting model with perceptual adversarial loss as MGANwriting.When generating images,the original GANwriting model cannot pay attention to certain image details,which results in distortion of the details of the generated image.The perceptual adversarial loss can control the generated image closer to the real image,and constrain the model from the perspective of the quality of image generation,so that the MGANwriting model can pay more attention to the part of image details.Experimental results show that compared with the original GANwriting model and the CycleGAN model,the MGANwriting model can generate Mongolian handwritten word images with higher quality,more diverse writing styles,and more accurate content,which effectively solves the problem of automatic generation of Mongolian handwritten samples.