Dual domain multi-scale feature enhancement network for light field denoising

Area of Science:

• Computer Vision
• Image Processing
• Optical Engineering

Background:

• Noise significantly degrades light field (LF) imaging quality, especially in handheld devices.
• Current LF denoising methods struggle with limited frequency domain analysis and integrating global/local features, leading to artifacts and poor structural consistency.

Purpose of the Study:

• To develop an advanced light field denoising method that overcomes limitations of existing approaches.
• To enhance the quality of LF images by effectively suppressing noise while preserving structural integrity.

Main Methods:

• Proposed a dual domain multi-scale network (DDMSN) integrating frequency-domain noise priors with pixel-domain texture and epipolar geometric awareness.
• Introduced a dual frequency collaborative enhancement (DFCE) strategy for differentiated frequency band processing.
• Implemented a global-local feature interaction (GLFI) mechanism within an encoder-decoder architecture for fused spatial-angular and EPI features.

Main Results:

• The DDMSN method demonstrated significant improvements in denoising performance compared to existing techniques.
• Achieved superior qualitative visual results and quantitative metrics on both synthetic and real-world noisy LF datasets.
• Successfully restored fine-grained frequency domain structures and preserved texture details and global consistency.

Conclusions:

• The proposed DDMSN effectively addresses the challenges in LF image denoising.
• The integration of dual domain processing and multi-scale analysis offers a robust solution for noise reduction in LF imaging.
• This approach significantly enhances the visual quality and structural fidelity of denoised light field images.