Research And Development Of Image Inpainting Algorithm Based On Global And Local Perceptual Adversarial Networks

Digital image inpainting technology is widely used in the computer field.Traditional digital image inpainting techniques are mostly based on structure and texture.If the damaged region is large or the semantic information is missing,the inpainting quality will be greatly reduced.In recent years,the generative adversarial network has shown strong ability to derive semantic information and capture image features,gradually has been attracted to the field of image inpainting.However,the current image inpainting methods based on the adversarial network still have problems of unstable training and insufficient inpainting quality.Therefore,in order to improve these two points,this paper researchs and develops an image inpainting algorithm based on global and local perceptual adversarial networks.This paper first introduces the principle of generating adversarial network and its improved algorithms,followly,proposes an image inpainting algorithm based on global and local perceptual adversarial networks,and then uses the TensorFlow computational library to implement the algorithm.Finally,the algorithm is evaluated by experiment.Among them,the following improvements have been made to the training stability and inpainting quality of the algorithm:(1)Based on the original model structure of the adversarial network,a local discriminative network is introduced to improve the inpainting quality of the demaged image.(2)In order to solve the difficulty in training of the network,and improve the overall quality of the inpainted image,Wasserstein Adversarial Loss is used to instead the original Generate Adversarial Loss.(3)The Perceptual Adversarial Loss is introduced,which calculates the difference between the inpainted image and the real image in the features of the hidden layers of the discriminative networks.By adversarial training,it helps the generative network to capture the differences between the output and the real image in the advanced features.In order to make the network first optimizes the pixel and structural similarity of the inpainted image,and then optimizes the advanced features of the inpainted image,the initial weights of the Perceptual Adversarial Loss are set to be a smaller value,and are increased with time.This paper uses the CelebA dataset and limit the damage type of the image to rectangular during training,then evaluates the algorithm by comparing with the CE and GLI algorithms.The results show that this algorithm is more stable in training,meanwhile,it has achieved better inpainting quality under the subjective evaluation indicators and objective evaluation indicators.