Chinese Character Inpainting Based On Improved Generative Adversarial Networks

As a carrier of Chinese culture,Chinese characters have played an important role in 5,000 years of history.However,a large number of ancient documents are damaged,it has seriously affected the study of Chinese ancient culture.With the development of image processing and pattern recognition,it is imperative to make the machine to repair damaged image automatically.Traditional image inpainting algorithms suffer from low accuracy and poor robustness.Deep learning algorithms,especially generative adversarial networks,as the hottest technology of artificial intelligence in recent years,have achieved remarkable results in the fields of image,voice,video,etc.In this thesis,Chinese character inpainting is modeled as an image generation problem,and the structure of the damaged part of the Chinese character is inferred using the pixel context information,and then the predicted part is filled in.It is proposed to use the context pixel information search for the input latent variable Z in the GAN network,and finally generate an image with a structure similar to the damaged image,then crop out the damaged part and backfill the damaged area.The main work of this thesis is as follows.(1)The end-to-end tablet denoising model is applied to the pre-processing by analysing the noise on the tablet,simulating the model of tablet noise,and then applying the generated noise to the clean tablet image to form the paired data required for training.(2)Conditional adversarial network(cGAN)is used to generate handwritten Chinese character.By adding conditional information to the GAN,the quality of Chinese character generation can be improved,and the problem of repairing errors(ambiguity)in later Chinese character repair can be solved by specifying conditional information.Comparative experiments with ACGAN have shown that the present method is superior in terms of both quality and stability of the results generated.(3)A pre-trained cGAN-based corrupted Chinese character inpainting method is proposed and implemented for the first time.Based on the fact that pre-trained cGAN can generate arbitrary Chinese characters,the search for the latent variable Z in the feature space of cGN by pixels in the undamaged region is used as input to cGAN,and the inference is made to obtain an output image that is very similar to the damged Chinese characters and backfilled.(4)The cGAN is further restricted by giving the conditional information of the character structure,so that it can fix the specified character to avoid errors and improve the correct rate of fixing.In conclusion,this thesis proposes a Chinese character inpainting method for the first time by using conditional adversarial network(cGAN)not only improves the quality of character generation in damaged areas,but also eliminates the occurrence of repair errors(ambiguities)during the repair process,and the algorithm is experimentally proven to have high accuracy and robustness.