Research On Ceramic Image Restoration Method Based On Deep Learning

Ceramic has survived for thousands of years and it has become a common material for people and princes and nobles in all dynasties with its good physicochemical properties.However,ceramics is supported by the inner lining of clay,so its biggest flaw is its fragility.Throughout the ages,too many ceramic objects have been too fragmented and broken to be studied accordingly,so this paper will use a series of image restoration technical methods to restore broken ceramic images.In the field of image restoration,traditional image restoration algorithms perform image restoration based on pixel-wise correlation and content similarity between images,such as sample-based methods and diffusion-based methods.Such methods can achieve relatively good results with small area defects and simple content images,but when the area of image defects is large or the texture structure of the image is complex,image restoration by traditional methods is ineffective.For these reasons,more and more researchers are focusing on deep learning-based image restoration methods,and deep learning for image restoration has become popular in the image research community because of its ability to produce higher quality restoration results.In view of this,a series of research and implementation of ceramic image restoration methods are carried out in this paper,using a combination of multiple deep learning image restoration models to achieve the purpose of ceramic image restoration,and comparing the effect of deep learning methods with traditional neural network restoration methods to explore the best ceramic image restoration solution.Deep learning methods for image restoration mainly include self-encoder network(AE)based methods and generative adversarial network(GAN)based methods.However,the images generated by the AE network are always unclear,and the generative adversarial network(GAN)approach is likely to generate uncontrollable images due to the high degree of freedom;the GAN can generate clear and high quality images because of the game between the generator and the discriminator,and the selfencoder network can obtain semantically rich images through the encoding and decoding process.Therefore,in this paper,a new image restoration model VAE/GAN is constructed by combining AE network and generative adversarial network.The method extracts the real sample feature information by encoder,and then lets the encoding approximate Gaussian distribution,and then dimensionally decouples the hidden variables obtained from the encoder for the purpose of discovering the association between dimensions.Factor VAE’s discriminator network is introduced in performing the hidden variable decoupling,and finally the generated reconstructed image and the image to be restored are fused with features to optimize the image restoration effect.The experiments show that the images restored by this fusion network are rich in texture and high in image quality by evaluating the generated images with different metrics.However,due to the long image training period,model collapse,gradient disappearance,and model non-convergence of GAN and its derived models.To address these problems,this paper proposes the use of DDPM model for ceramic image restoration.Through the process of noise addition and denoising,the neural network model can then perform real image generation for noisy images,and accordingly,the image itself defects will be repaired by the DDPM model during the image generation process.Experiments show that the model can effectively repair ceramic images,and the repaired images have clear textures.Finally,this paper compares the restoration effects of these types of models to provide a reference for ceramic image restoration research.