Research On Image Inpainting Algorithm Based On Generative Adversarial Network

As an important research direction in the field of computer vision,image inpainting has important applications in the fields of image enhancement,PS,3D reconstruction and face recognition.In recent years,the development of deep learning has led to the emergence of a large number of research results related to image inpainting.However,the existing deep learning-based methods still have many shortcomings in the field of image inpainting.In response to these problems,this paper improves the existing network,and the specific work is as follows:Firstly,aiming at the problem that the content and structure of the image are blurred by repairing the image under the condition of large-area defect,the gridding effect caused by the convolution of stacking dilated convolution in the current model is analyzed,and the multiscale dilated convolution composed of multiple parallel dilated convolution with different dilation rate is proposed.The encoder-decoder-based method replaces the middle layer of the existing network and adds several multi-scale dilated convolution modules,which enhances the network’s ability to extract information breadth and precision at multiple levels.The problem of gradient disappearance caused by distribution shift in training using instance normalization network is analyzed.The method of region normalization is introduced,which replaces the original batch normalization operation in the network normalization layer.Then,in order to generate finer textures and clear structures,generative adversarial networks are used.The traditional contextual attention mechanism is improved,the feature block size category in the original mechanism is increased,and the reconstructed feature map is fused by the squeeze-and-excitation module,so that the network can reconstruct the correct features of the defect area during feature matching,resulting that the details of the image can be repaired richer.Based on the U-Net network structure and using the improved contextual attention mechanism,a coarse-to-fine two-stage generative network is designed.A method using Gaussian filtering is proposed to smooth the label values at the boundary,and improve the adversarial loss function.