Image Inpainting Based On Differentiable Neural Architecture Search

Image inpainting is one of the research hotspots in the field of image processing.Although image inpainting has been around for many years,it has gained even more popularity because of the recent development in image processing techniques.Image inpainting mainly aims at repairing the missing area or damaged image,so that the repaired image can be as close to the original image as possible.Because of the selflearning ability of neural network,the image inpainting method based on deep learning has achieved good results.However,the current neural network architectures are mostly based on manual design.Normally,a good neural network architecture will take too much time and resource,so automatic deep learning becomes particularly important for specific problems.Neural architecture search is one of the main methods of automatic deep learning.By using different search strategies,we can get a more suitable neural network architecture for the current problem.In order to extend neural architecture search technology to image inpainting,this paper proposes an image inpainting algorithm based on differentiable neural architecture search.First of all,according to the structure of encoder-decoder,a new search cell called upsample cell is proposed for image up sampling operation.And it appears in pairs with the down sampling operation called reduction cell.In order to save the time of architecture searching and improve the efficiency of computing,the depth of network is increased progressively by increasing cell symmetrically.Symmetrical increase of reduction cell and upsample cell,or symmetrical increase of normal cell.With the development of the search process,the depth of the network increases progressively.The search process is divided into several stages,and two cell structures are added in each stage.The DARTS algorithm based on differentiable neural architecture search is applied to image inpainting problem,and it is used as a baseline.Then,the improved differential neural architecture search proposed in this paper is also applied to image inpainting.Experiment are carried on CIFAR-10 dataset and SVHN dataset respectively.The experimental results show that the image inpainting algorithm based on differential neural architecture search proposed in this paper not only has a great improvement in the inpainting effect,but also has achieved state-of-the-art performance on search time.The PSNR index on CIFAR-10 dataset is 32.8952 db,and the search time is 0.4 GPU days.