A Novel Approach for Image Inpainting using GAN

A technique to fill in the missing areas in an image by obtaining information within the image is known as image inpainting. Recent advances in image inpainting using deep learning have demonstrated impressive performance in reconstructing images with missing regions. However, they fail to generate missing regions in images where faces are unaligned. When an image features a person wearing glasses and if some areas corresponding to the glasses are masked, the state-of-the-art models often struggle to accurately inpaint these regions, resulting in outputs that lack semantic accuracy. Hence, to overcome the mentioned limitations, this paper introduces a novel GAN-based model for image inpainting that produces complex and detailed patterns by capturing the distant visual information using contextual transformation blocks. The model undergoes a comprehensive evaluation, using the CelebA-HQ dataset for training and the FF-HQ dataset for testing. The experimental results reveal that our approach provides significant reduction up to 41.16% in LPIPS, as well as improvement up to 12.28% in PSNR and 4.2% in SSIM, compared to the state-of-the-art techniques.