Research On Multi-stage Image Inpainting Algorithm Based On Generative Adversarial Network

Image inpainting,which can complete missing information based on existing information,is considered as an image-processing technology with significant research and application potential in military security,biomedicine,and other fields.However,affected by the size of missing areas and the referability of existing areas,the inpainting results may have distorted structures and blurred textures,which limits the application of image inpainting technology in high-accuracy fields.Based on the review of state-of-theart image inpainting technologies,this thesis proposes two effective algorithms to improve the accuracy of image inpainting,which considers the context inconsistency problem and texture blurring problem.The research work of this thesis can be summarized as follows:In a general image inpainting task without guidance information,a multi-stage highlow-frequency-fusion image inpainting algorithm is proposed based on the generative adversarial network(GAN).The algorithm focuses on the target inconsistency problem between different loss functions,and consists of high-low-frequency prior-information generation module(HLF-GM)and high-low-frequency fusion module(HLF-FM).In HLF-GM,the module generates different high-low frequency prior information according to different loss functions.Through the dedicated training of high-low frequency prior information,the algorithm can reduce the impact of inconsistent targets of different loss functions.In low-frequency domain,a smooth low-frequency image is generated by the reconstruction loss function.In the high-frequency domain,a high-frequency image is generated by the adversarial loss function and the perceptive loss function;In HLF-FM,the generated prior information is fused and utilized to refine the inpainting results.Focal frequency loss is utilized to guarantee the generation of low-priority and difficult-togenerate frequency information,that can avoid the loss of frequency information.Through the pixel-by-pixel discrimination of the U-net discriminator,the refinement of local texture information is realized.Moreover,by embedding the Fast Fourier Convolutional Block(FFC-Block)in U-net in both HLF-GM and HLF-FM,the algorithm can utilize global and local context information more effectively.The accuracy of the algorithm is proved by experiments with commonly used image inpainting datasets.In an image inpainting task with guidance information,this thesis proposes a twostage guiding algorithm based on GAN.In the first stage,a structure image is reconstructed.A method that injects guidance information in the skip-connection layer is adopted to avoid losing the effectiveness of guidance information in convolution and regularization processes.Moreover,the attention mechanism is applied to allocate the weight of the feature map.In the second stage,a pyramid-structure U-net guides the generation of low-level texture information by high-level structure information.A texture contrastive loss function is designed to reduce the feature distance between the original image with abundant texture information and the inpainting results while increasing the feature distance between the structure image and the inpainting results.The texture contrastive loss function can improve the authenticity and richness of texture details,and it can solve the problem that the guidance information is too sparse to effectively guide the inpainting of texture information.Through the experiments with commonly used image inpainting datasets,it is proved that the proposed algorithm can improve the image inpainting effects.The image inpainting results can be controlled by guidance information,and its texture information can be more authentic by the utilization of the texture contrastive loss function.