INTELLIGENCE MULTI-FOCUS IMAGE FUSION

Abstract

In this paper, we survey the most recent advances in Multi-Focus Image Fusion (MFIF), with special focus on breakthroughs which have been achieved under deep leaning (DL) paradigms. Conventional MFIF methods (either in spatial or frequency domain) are problematic concerning sensitivity to mis-registration, noise and artifacts. The emergence of DL techniques, such as Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), Auto encoders and Transformers, has significantly improved quality of fusion by learning complex feature representation and optimal fusion strategies directly from data. Transformer-based methods, including Swin Fusion, have achieved better performance in modelling long-range relations and highlighting informative characteristics, leading to competitive SSIM and compared with previous works. However, these models struggle with technical issues in practice due to high computational cost preventing real-time execution on embedded systems and their vulnerability to real-world artifacts such as noise and mis registration. The paper emphasizes Utility of evaluation metrics (SSIM, EN, SF and MI) for measuring fusion quality multi-metric indices are more believable for quality measurement. The major research gaps include larger (real world) datasets, domain-independent robust models, lightweight architectures for real-time applications and interpretability of DL models. Outlook meanwhile highlights hybrid models involving DL with classical approaches such as graph theory, and the construction of robust and computationally efficient networks. In summary, this review highlights that despite the advancement brought by deep learning to MFIF, addressing deployment challenges and improving robustness is crucial for practical usage in areas like surveillance, microscopy and medical imaging.