A Joint Loss based Deep Learning Model for Enhanced Image Inpainting

Research on Image inpainting has attracted growing attention as image data and applications become commonplace. Nevertheless, it is still a challenge to repair or restore missing parts of electronic images to its original state. In this paper, we develop a new deep learning based model by combining the two mainstream generative models: Auto-encoders and Generative Adversarial Networks to improve the effect of image inpainting. Specifically we design a novel joint loss function which take into consideration reconstruction loss in a Generator and adversarial loss in a Discriminator to improve the repaired effect of texture features and semantic information. The model and joint loss function were experimented and evaluated using the public face data set celebA. Initial results showed that our model and function has significantly improved the effect of repairment in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics.