DCGAN-based Algorithm For Inpainting The Defect In The Image

As the cost of visual information acquisition equipment continues to decrease,images,as an important carrier of visual information,have led to an increase in the average daily image processing size of the global Internet.In the process of image generation,transmission and downloading,there is a tendency for images to be damaged and contaminated,requiring restoration work to be carried out on missing areas.Among the images to be restored,a high percentage of images are of buildings.Existing image restoration techniques are widely used in architectural design,industrial manufacturing,film and television production and other industries.The basic principle of image restoration is to use the semantic information of an image in a known area,and then filter,copy and expand the missing area to make a logical addition to the original image.Traditional image restoration algorithms are suitable for restoring images with simple textures,a few missing areas and monotonous backgrounds,while deep learning-based image restoration algorithms are suitable for restoring images with many details,a large number of missing areas and overall complexity,but current deep learning-based image restoration algorithms exhibit poor restoration results,unstable model training and low applicability of the algorithms.Among such algorithms,the generative adversarial network model is more desirable,and the model’s generator is more realistic in the restored image,enhancing the development of the image restoration field.In this thesis,we use the DCGAN model as the base algorithm,mainly improving on the generator and discriminator,with the main restoration object being building images.The generator model incorporates a self-attentive mechanism,the discriminator model sets up a global discriminator and a local discriminator,and the image input part sets up an Edge-GAN-based complete texture complementation of the image.The overall model implementation process is presented throughout the thesis.Firstly,validation experiments are done for each model part to demonstrate the feasibility of repairing defective images of buildings.Then ablation experiments are done on each model to compare the impact of each part on the overall model.Finally,experiments are conducted on the overall model with different data sets to demonstrate the enhancement of the restoration of defective images of buildings.The experiments are set up with different sizes of defective masked areas,and the post-repair evaluation criteria are a combination of subjective image comparison and objective parameter evaluation.The subjective evaluation is the post-repair effect map of different algorithms,which indicates the respondents’ evaluation of the visual perception of the repaired images,and the objective parameters are PSNR and SSIM,which are used to compare the restoration results of different algorithms under the same hardware environment horizontally.The data from the restoration results and objective parameters show that the complete algorithm in this thesis has improved restoration results in the defective areas of buildings compared to other algorithms.In the case of a medium area of missing building images,the model achieves the best restoration results compared to other algorithms.There is an improvement in the structural similarity index(SSIM)metric and a 20% improvement in the peak signal-to-noise ratio(PSNR)metric.In scenes where objects obscure buildings,the restoration of details such as building textures and edges performs better than other restoration algorithms,and the restoration algorithm in this thesis also proves experimentally suitable for application to image restoration of self-built building datasets,which has practical translation value.