Image Restoration Algorithm Based On Neural Network And Attention Mechanism

The image inpainting dates back to the Renaissance,which is used to fill the missing parts of the paintings at that time.The technology can be widely used in many fields such as medical,military,film,education and so on,which is attracting more and more scholars to invest in research.However,traditional image inpainting models cannot actively find the location of the area which needs to be repaired,and cannot repair fully dynamic images.There is no unified model that can deal with the problem of random damage in both time and space.In recent years,the rapid development and outstanding performance of deep learning in the image processing direction have given researchers new ideas and inspirations.Firstly,the article introduces autoencoders,generative adversarial networks,deep separable convolutions,attention and progressive growth mechanisms in the field of deep learning.And article focuses on good characteristics in the field of image inpainting,which studies both the respective implementation mechanisms and internal correlation.Then,the paper presents an improved image inpainting model,which is combined with the advantages of autoencoder and generative adversarial network.To enhance the ability of extracting local details,the model introduces conditional information and attention mechanism.The progressively increasing training mechanism is introduced to increase the resolution of generated images steadily.In order to control the generation of features at all levels of the image,the definition formula of hidden vector of autoencoder in the generator part is improved,and new random noise is added to ensure the diversity of the details of the generated image step by step.Through the longitudinal comparison of the model in the article under different configurations,and the horizontal comparison between the model in the article and the other three mainstream image restoration models CVAE,DCGAN,and WGAN,mainly from the visual effects and numerical evaluation indicators,the effectiveness of each improved method to improve the model performance and the superiority ofthe image restoration performance of the neural network proposed in the paper are verified.Finally,to reduce the amount of network parameters and calculations and thus increase the number of running frames of the model on hardware devices,such as replacing all traditional convolutions with deep separable convolutional residual structures,weight initialization,weight pruning and half-precision transformation optimize the model.The experimental results show that the improved algorithm in the paper has better inpainting performance.