Review of Deep Learning-Based Image Inpainting Techniques

The deep learning-based image inpainting models discussed in this review are critical image processing techniques for filling in missing or removed regions in static planar images, and they have been extensively researched and applied. However, due to the rapid advancements in deep learning technologies, existing review studies exhibit shortcomings: in 1) providing a comprehensive and systematic classification of strategies for these models; 2) analyzing loss functions and evaluation metrics; and 3) organizing and categorizing data and application domains. To address these three aspects, this paper provides a systematic literature review of deep learning-based inpainting models. First, we rigorously classify the strategies used in image inpainting networks into four categories: layer-based, connection-based, multi-network, and multi-modal inpainting strategies. This classification is determined by the composition of the models, ensuring it comprehensively covers the vast majority of inpainting network strategies. Second, to facilitate a more effective study and comparison of loss functions and evaluation metrics, we categorize them into pixel-based, feature-based, and model-based approaches, conducting a thorough analysis of each. Finally, we categorize image inpainting applications into three main types according to dataset characteristics: natural image inpainting, detection image inpainting, and artistic image inpainting, offering insights into each specific domain.