Rethinking Few-Shot Image Fusion: Granular Ball Priors Enable General-Purpose Deep Fusion
Researchers' new method trains on just ten image pairs, achieving state-of-the-art fusion with a lightweight network.
A research team has published a paper titled 'Rethinking Few-Shot Image Fusion: Granular Ball Priors Enable General-Purpose Deep Fusion,' introducing a breakthrough method that drastically reduces the data needed for AI-powered image fusion. The core innovation is the concept of 'incomplete priors,' which formally describes and estimates the confidence of handcrafted rules guiding the fusion process. These priors are coupled with a neural network via a sample-level adaptive loss function, allowing the model to learn and re-infer fusion rules that closely approximate real-world conditions.
To generate these priors, the team developed the Granular Ball Pixel Computation (GBPC) algorithm, rooted in granular computing principles. GBPC models image pixels as information units, performing fine-grained pixel weight estimation while statistically evaluating prior reliability at a coarse-grained level. This dual-scale approach enables the system to perceive discrepancies between different image modalities (like infrared and visible light) and adapt accordingly. The result is a method that, in extensive experiments, achieved superior performance in visual quality and model compactness, even when trained on patches from just ten image pairs.
- Introduces 'incomplete priors' and a GBPC algorithm for fine-grained pixel control and prior confidence estimation.
- Trains a lightweight neural network effectively with only 10 image pairs, a >90% reduction in required data.
- Demonstrates state-of-the-art performance across diverse fusion tasks including infrared-visible and medical imaging.
Why It Matters
Enables high-quality AI image fusion for medical, surveillance, and photography with minimal training data, lowering barriers to deployment.