Research On Image Completion Algorithm Based On Generative Adversarial Network

Image completion technology is a technology that uses reasonable pixel values to fill the missing areas in the image.It’s difficult for traditional image completion methods to achieve reasonable results when the damaged area has complex structure and contains a large amount of semantic information.In recent years,with the development of deep learning,the application of generative adversarial network in image completion has become a hot research direction,which is widely used in all walks of life.Through game learning and mutual optimization between generation and discrimination networks,the generative adversarial network makes the network generate samples close to the real distribution,which makes the image completion technology of generative adversarial network very consistent.Therefore,the image completion technology based on generative adversarial network has important research significance.With the development and in-depth research of image completion model,from the completion of damaged areas with fixed rectangular shape to the completion of damaged areas with arbitrary shape,the practicability of image completion algorithm continues to improve,and the results of image completion are closer and closer to the real image.The requirements for image completion are not limited to the results similar to the original image.Instead,the technology has been developed to generate new images with consistent semantic style and covering more scenes.However,there are still some problems and room for improvement.For example,for the completion of regular damaged areas with large size,the existing methods have limited receptive field and insufficient semantic information acquisition resulting in the fuzzy effect of the details of the completed image,and the global consistency with the original drawing is poor.In addition,for the completion of irregular damaged areas,when there is a mixture of foreground and background in the image damaged areas,the existing algorithms extract semantic information by interfering with invalid pixels in the damaged areas resulting in the error distortion of the structure information,semantic ambiguity and fuzzy detail texture in the completion results.In order to solve the above problems,two image completion models based on generative adversarial network theory and existing completion models are proposed.The main research contents of this paper are as follows:(1)Aiming at solving the problems of fuzziness,poor consistency between the completed area and the original image,long training time and low efficiency in the image completion of regular damaged image with large size,this paper proposes a regular damaged image completion algorithm with improved generation network structure.The algorithm’s generation network adopts self coding structure,and the middle layer design uses JPU(joint pyramid upsampling)module to replace the expanded convolution layer connected by a large number of residuals to enrich the scale of semantic feature extraction and improve the speed of model training.The identification network adopts the global and local double discrimination network structure,which not only ensures the global consistency between the completed area and the original image,but also enriches the texture details and improves the clarity of the repaired area.The proposed model is compared with CE and GLCIC models in the image completion of regular damaged image with large size.The experimental results show that the image completion effect of this model is better on places 2 and Paris Streetview data sets,and the objective evaluation indexes PSNR and SSIM are better than other algorithms.(2)Aiming at solving the problems of wrong distortion of structural information,semantic ambiguity and fuzzy detail texture in the image completion of irregular damaged areas of arbitrary shape,this paper proposes an irregular damaged image completion algorithm based on edge prediction.The algorithm uses two-stage to generate the generative adversarial network structure.In the first stage,the edge structure information of the missing areas of the image is predicted and completed to generate the edge knot In the second stage,the image completion network takes the predicted edge map as a priori information to complete the damaged area of the image,so as to make the structure information of the completed area more accurate.The attention based multi-scale expansion convolution fusion structure MSA(Multi-Scale Attention)module is added to the image completion network to extract the feature semantic information of the image.The proposed model is compared with Ca and Foreground-aware models in the image completion of irregular damaged areas of arbitrary shape.The experimental results show that the visual effect of image completion of this model is more real on places 2 data set,and the objective evaluation indexes PSNR and SSIM are better than other algorithms.