Research On A GAN-based Denoising Method For Chinese Stele And Rubbing Calligraphic Image

With the massive development of digital library platforms,Chinese stele calligraphy images show significant historical and artistic value.However,due to natural weathering corrosion or man-made factors,the ancient stele had been seriously damaged.So,the collected images were inevitably polluted by noise.This will bring great difficulties and regrets to people studying calligraphy.Therefore,it is undoubtedly of great significance to the inheritance and development of calligraphy to study how to denoise the Chinses calligraphy images.In recent years,denoising algorithms based on deep learning have sprung up.However,there are still many difficulties for stele calligraphy image denoising task.For example,the dataset of stele calligraphy images is not sufficient;the denoising methods cannot effectively deal with various types of noise in the stele calligraphy images;the denoising methods cannot remove the noise while ensuring that the font strokes are not damaged.So,this thesis conducts a detailed study on the denoising problem in stele and rubbing calligraphy images.This thesis will conduct an in-depth study on deep learningbased denoising method by combining generative adversarial networks,dense connection mechanisms,attention mechanisms,and dilated convolutions.The main research work of this thesis are as follows:In order to solve the problem of few data pairs in the stele and rubbing calligraphy image dataset,an offline-based data augmentation method is proposed.The method firstly extracts the noise image based on the image noise model,given the known noise image and its corresponding clean image.Then,a new noise image is obtained by various basic transformations on the noise image.Finally,the new noise image is added to the clean image or superimposed on the existing noise image,and then a large number of image data pairs are constructed to complete the task of data augmentation.In order to solve the task of denoising stele and rubbing calligraphy images,a denoising model of stele and rubbing calligraphy images based on generative adversarial network is proposed.First,for the problem that it is difficult to fully utilize the features extracted by each layer in the network,a densely connected block is introduced in the generator.Densely connected block can fully utilize the information in each convolutional layer.And the idea of Patch GAN is used in the discriminator to train the discriminator using small block images.The advantage is that it cannot only reduce the parameters of the network,but also maintain the high resolution and local texture details of the generated images in the image generation task.In the dataset of stele calligraphy images with different noise types,the experimental results show that this method has better performance in dealing with point noise and scratch noise.In the denoising method designed for the above denoising tasks,there are problems that some noises in the denoised image are not completely eliminated and font strokes are lost.In this thesis,we consider a combination of attention mechanism and autoencoder,and design a denoising method for stele and rubbing calligraphy images based on attention mechanism adversarial generation network.This network is a two-stage denoising method.Firstly,the attention mechanism is used to find the noise regions in the noisy image by the recurrent network.Then,the denoising autoencoder is guided by a noise map for denoising.The denoising autoencoder designed in this thesis uses dilated convolution instead of ordinary convolution to improve the receptive field of the network,and it’s verified from experiments that it can improve the denoising performance of the method.The experimental results show that the attention mechanism is effective and better than other end-to-end methods in denoising performance.The method can effectively preserve the stroke details in the original image font after denoising.And it can show excellent performance in dealing with stele calligraphy images of various types of noise.