Abstract

Multi-focus image fusion (MFIF) is an image enhancement technique that investigates how to obtain a fully focused image from multiple defocus images, providing fundamental services for computer vision fields, such as image recognition, 3D reconstruction, medical diagnosis, computational photography, etc. In recent years, the deep learning-based MFIF has broken through the limitations of traditional MFIF and achieved superior fusion results. Existing reviews on MFIF mainly classify and describe the techniques at the technical level, such as supervised/unsupervised learning methods and network types, but lack discussion and analysis on problem scenarios. Therefore, based on the problem scenarios of MFIF, this paper categorizes deep learning-based MFIF research into six types: MFIF with lightweight networks, MFIF for artifacts and defocus spread effects, MFIF for information preservation, MFIF with unified fusion networks, MFIF on addressing suboptimal initial decision map and MFIF in challenging environments. Furthermore, this paper summarizes commonly used synthesis datasets and real datasets for MFIF, and statistically describes the main evaluation metrics. Finally, this paper analyzes the shortcomings of existing algorithms, and identifies potential future research directions, aiming to provide objective reviews for MFIF researchers focusing on different problem scenarios.