Research On Image Inpainting Algorithm Based On Parallel Convolution And Multi-Information Fution

Aiming at the defects of unreasonable repair structure and fuzzy generation details when repairing a large area of random missing area with the existing deep network,this paper proposes two improved repair models: two-stage adversarial network image inpainting based on parallel convolution and single-stage adversarial network image inpainting based on multi-scale information fusion.The main contributions are as follows:Aiming at the problems of discontinuous structural information and missing semantic information in existing algorithms when repairing large-area regular masks,a two-stage inpainting model based on parallel convolution is proposed,which consists of a coarse inpainting network composed of parallel convolutions and a fine inpainting network that fuses residual connections and attention mechanisms.Firstly,the masked image is fed into the coarse inpainting network to extract the image structure information,and generate the coarse prediction image,as well as calculate the pixel gap between the coarse prediction image and the original image to optimize the coarse inpainting network.Subsequently,the coarse prediction image is inputted into the fine inpainting network,and an attention mechanism is introduced to the fine inpainting network,paying special attention to the image details,so that the content generated by the network is clearer.Finally,the global and local discriminators and pre-trained VGG network are applied to extract image information,and the loss function is introduced to optimize the network reasonably,improve the network repair ability and make the repaired area more consistent with the original area information.The proposed algorithm is trained and tested using public data sets,and compared with the images generated by the latest algorithm.The experimental results demonstrate that the repair information in this method is more reasonable in repair large-regular missing areas with complex textures and structures,improve the authenticity and integrity of image details,semantics and structure,illustrating superior performance over the classical comparison algorithm in terms of peak signal-to-noise ratio and L2 distance.Aiming at the problems of structural distortion and texture blur in existing algorithms when repairing large irregular missing area,a inpainting model based on multi-scale information fusion is proposed.The model consists of a basic generator and a discriminator,The generator designs a multi-scale information fusion block to be inserted between the encoder and the decoder,a skip connection is introduced,and designs a code information fusion block for the connection line to better fuse the information generated by the encoder and the information generated by the decoder.Firstly,the distorted image structure information is extracted by the encoder to predict the missing information.Subsequently,the multi-scale information of the image is extracted by the multi-scale information fusion block.Finally the image is reconstructed by the decoder.Reconstruction loss and adversarial loss are jointly implemented to calculate the gap between the reconstructed image and the real image to optimize the generator to make the content generated by the network clearer.The proposed algorithm is trained and tested using public data sets,and compared to the images generated by other methods.The experimental results demonstrate that the method can repair finer results when there is large and randomly irregular semantic loss,which the peak signal noise ratio,frechet distance and other indices are superior over the classical comparison algorithms.