Research On Digital Image Restoration Algorithm Based On Auto Encoder And GAN

Digital image restoration technology refers to the use of information in known areas in digital images to automatically repair the contents of defective areas in images.Digital image restoration algorithms based on deep learning occupy a dominant position in the field of image restoration due to their powerful data fitting capabilities and remarkable phased results.However,when the image texture and structure are relatively complex and the defect area is large,the existing deep learning-based image restoration often has problems such as lack of fine texture,local structure distortion,or texture and structure inconsistency,and the repair effect is not ideal.In response to the above problems,this thesis takes the generative adversarial network as the backbone network,and improves the network by integrating U-net network,attention mechanism and structural auxiliary branches,which effectively improves the quality of image restoration.The main work of this thesis is as follows:1.A post-position attention digital image inpainting algorithm is proposed.The main network of the algorithm is generative adversarial network,integrates U-net network as the generator,and uses skip connections to supplement low-level image features such as texture for decoding feature maps.By replacing ordinary convolution with partial convolution,the performance of the network in repairing irregular defect images is improved.A post-position attention module is added to extract relevant features from known areas far away from the defect area,which enhances the consistency between the local features of the image and the overall image,and makes the repaired image have more detailed textures.Using a Markov discriminator,by evaluating the generated image in blocks,it achieves a more reasonable loss representation than traditional global discriminators,and further improves the quality of image inpainting.2.A digital image inpainting algorithm with a structure-assisted branch is proposed.On the basis of post-position attention digital image inpainting algorithm,a double-branch network is constructed by using the structural information of the image edge map and adding structural auxiliary branches.Add a feature fusion module with information selection function,integrate the structure feature map and texture feature map,make the feature information from the two branches complement each other,and enhance the consistency of the repair image structure and texture.A learnable parameter is introduced into the post-position attention module to improve the module’s ability to extract long-distance feature information.At the same time,a structural Markov discriminator is added on the basis of the original Markov discriminator,and the two Markov discriminators jointly play a game with the generator,forcing the generator to generate a more realistic repair image.