Research Of Face Image Inpainting Algorithm Based On Edge Guidance And Multi-Scale Dense Convolution

In recent years,due to the impact of human,environment,and other factors on the transmission and acquisition of images,resulting in the loss of image information,it is necessary to repair the images and enhance their visual effects.Traditional image inpainting methods,since mainly inferring image missing regions based on the existing texture information and structural information,are unable to achieve the desired results for complex face images.On the other hand,the application of deep learning has greatly improved performance of image inpainting.However,for face image inpainting with large irregular missing areas,it may induce image structural distortion,detail blurring,poor fusion effect,and inaccurate feature information restoration and so on.Therefore,in order to solve these problems,the main work is as follows:1)A three-stage model based on structure prediction and image completion is proposed to repair face image with large irregular missing areas.In the first stage,an encoder-decoder network with dilated convolutions is used to perform initial repair of missing parts via contextual information from surrounding image features.In the second stage,the output at the first stage is fed into an encoder-decoder network based on a selfattention mechanism to predict the edge structure of the missing regions.Finally,the outputs of the first two stages are passed to a refined inpainting network to guide the repair process using the improved U-net architecture.In simulation,we use the classical algorithms as benchmarks to verify the performance of proposed approach on a public dataset.The experimental results show that the proposed method is superior to existing methods in subjective visual and objective evaluation.2)A progressive face image inpainting method based on multi-scale dense convolution is proposed to complete the segmentation map reconstruction and detail repair of images.In the first stage,a gated convolution-based coding-decoding network is used to reconstruct the missing area of the segmentation map so as to guide the subsequent image inpainting.In the second stage,the reconstructed segmentation map in the first stage and the original damaged image is fed into a multi-scale dense convolution-based detail repair network to restore the texture information of the missing area.The proposed algorithm is compared with existing classical algorithms on a public dataset,and the experimental results show that the proposed method is superior to existing methods in subjective visual and objective evaluation.