Incrementally Adapting Pretrained Model Using Network Prior for Multi-Focus Image Fusion.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Multi-focus image fusion aims to create a single all-in-focus image from multiple source images with varying focal lengths.
• Existing supervised methods suffer from domain shift on real-world data, while unsupervised methods lack general defocus blur knowledge.

Purpose of the Study:

• To develop a novel multi-focus image fusion model that overcomes limitations of current supervised and unsupervised approaches.
• To improve the performance of image fusion by integrating general knowledge and sample-specific priors.

Main Methods:

• Proposed the Incremental Network Prior Adaptation (INPA) framework.
• Leveraged a supervised pretrained backbone for general knowledge.
• Integrated extrinsic priors optimized on specific testing samples into a smaller prior network.

Main Results:

• The INPA framework effectively integrates features from strong baselines into a compact prior network.
• Evaluated on synthetic and real-world datasets (Lytro, MFI-WHU, Real-MFF).
• Demonstrated superior performance compared to existing supervised and unsupervised multi-focus image fusion methods.

Conclusions:

• The proposed INPA framework offers a robust solution for multi-focus image fusion.
• This hybrid approach successfully balances generalizability and specificity for improved fusion quality.
• The method shows significant advancements in generating all-in-focus images from multi-focus sources.