FDSNet: Frequency-Decoupled Stack Fusion Network for Light Field All-in-Focus Image Generation

Area of Science:

• Computational Imaging
• Computer Vision
• Image Processing

Background:

• All-in-focus (AIF) images are vital for light field (LF) measurement but typically require depth priors.
• Existing multi-focus image fusion (MFIF) methods struggle with LF data stacks due to error accumulation.

Purpose of the Study:

• To develop a depth-free method for high-precision AIF image generation from LF data.
• To overcome limitations of existing MFIF methods in handling LF image stacks.

Main Methods:

• Proposing the Frequency-Decoupled Stack Fusion Network (FDSNet).
• Utilizing a spatial-frequency joint feature extraction module to decouple high- and low-frequency components.
• Employing a dual-stage cross-attention fusion module with a coarse-to-fine strategy for artifact suppression and edge fidelity.

Main Results:

• FDSNet achieves superior visual quality and quantitative performance on synthetic and real LF datasets.
• The network demonstrates robustness under low-light and noisy conditions.
• FDSNet effectively avoids error accumulation and computational redundancy inherent in iterative fusion.

Conclusions:

• FDSNet offers excellent fusion capability, excelling in image clarity, detail preservation, noise resistance, and generalization.
• The proposed method provides a significant advancement for depth-free LF AIF image generation.
• FDSNet outperforms state-of-the-art methods in LF image fusion tasks.