Rethinking Multi-Focus Image Fusion: An Input Space Optimization View.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Multi-focus image fusion (MFIF) integrates images with varying focal depths to overcome partial focus limitations.
• Existing MFIF methods often require complex loss functions or extensive synthetic datasets, limiting their practical application.

Purpose of the Study:

• To propose a novel, cost-effective approach for multi-focus image fusion (MFIF) by optimizing the input space.
• To develop a framework that leverages intermediate features from pre-trained, non-MFIF networks for improved fusion quality.

Main Methods:

• A cascaded framework was designed, incorporating two feature extractors (image deblurring and salient object detection networks).
• A Feature Distillation and Fusion Module (FDFM) was proposed as a training-free adapter to create an MFIF-suitable input space.
• An isometric domain transformation, extending manifold theory, was developed to handle high-dimensional feature maps within FDFM.

Main Results:

• The proposed model demonstrated superior performance compared to 13 state-of-the-art methods across six benchmark datasets in both qualitative and quantitative analyses.
• The optimized input space significantly improved the performance of various MFIF models without requiring additional training or data.

Conclusions:

• The novel input space optimization strategy offers a significant advancement in multi-focus image fusion.
• This method provides a cost-effective and efficient solution for enhancing MFIF, applicable to existing models.