A Review of the Foundations, Developments and Prospects of Image Inpainting Based on Deep Learning

Abstract

Images are visual medium used to convey information, express opinions, record events or display artistic creations and are widely used across diverse fields. However, they may suffer from various defects arising from poor storage methods, inadequate devices, inappropriate techniques or human damage. In the past, traditional methods were primarily used for image inpainting. Through the employment of deep models, image inpainting based on deep learning techniques can boost accuracy in preserving image texture and structure. Despite the prominence of image inpainting in the field of computer vision, there remains a dearth of comprehensive review works. This study provides a comprehensive analysis of sophisticated image inpainting methodologies based on deep learning models, considering common datasets and assessment criteria. In particular, the examined models comprise autoencoder-based models, U-net-based models, generative adversarial network-based models, and transformer-based models. Additionally, potential pathways for future research are also discussed.