Deep learning for image inpainting: A taxonomy, review, and future directions

Image inpainting reconstructs missing regions in images using contextual information, with applications spanning digital restoration, medical imaging, and content editing. Advances in deep learning have shifted from simple convolutional methods to sophisticated deep generative models, including Variational Auto-Encoders, Generative Adversarial Networks, attention mechanism, and Diffusion Models. To help researchers gain a clear understanding of image inpainting theories and facilitate the initiation of their studies, this paper presents a comprehensive review of research progress made over the past 5 years. The inpainting methods are categorized into monistic and pluralistic approaches according to task type, with further subdivisions by inpainting strategies and deep generative models applied. The paper also summarizes common datasets, loss functions, and evaluation metrics in the field. Additionally, representative monistic and pluralistic methods are evaluated qualitatively and quantitatively on the Places2 dataset. Finally, the future research directions alongside emerging trends are discussed.