FMDL-Net: Fourier modulation and dynamic mixing for light field image super-resolution

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Light field imaging faces a spatial-angular resolution trade-off.
• Current deep learning methods like CNNs and Vision Transformers have limitations in capturing long-range dependencies and preserving high-frequency details.

Purpose of the Study:

• To develop an advanced deep learning model for light field image super-resolution (LFISR).
• To address the limitations of existing methods in capturing global dependencies and preserving texture details.

Main Methods:

• Proposed FMDL-Net, a disentangled network integrating Fourier Modulated Attention (FMA) and Dynamic Mixing Layers (DML).
• FMA utilizes Fast Fourier Transform for efficient global receptive fields and geometric consistency.
• DML employs content-adaptive dynamic convolutions for texture preservation and disparity handling.

Main Results:

• FMDL-Net demonstrated superior performance over state-of-the-art methods on five public datasets.
• Achieved higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) scores.
• Required fewer parameters and lower computational cost compared to existing approaches.

Conclusions:

• FMDL-Net effectively overcomes the spatial-angular resolution trade-off in LFISR.
• The proposed FMA and DML modules enhance the model's ability to capture global context and preserve intricate details.
• FMDL-Net offers a computationally efficient and high-performing solution for LFISR.