Robust light field angular super-resolution via multi-dimensional feature fusion and attention-guided refinement

Area of Science:

• Computer Vision
• Image Processing
• Computational Imaging

Background:

• Angular super-resolution (ASR) is crucial for reconstructing dense light fields (LF) from sparse views.
• Existing ASR methods face challenges in maintaining consistency with occlusions and large disparities.
• Robust reconstruction is needed for complex real-world scenarios.

Purpose of the Study:

• To propose a robust attention-guided multi-dimensional feature fusion network (LFAMF) for light field angular reconstruction.
• To improve the preservation of structural integrity and angular consistency in challenging imaging conditions.
• To enhance the performance of angular super-resolution in light field imaging.

Main Methods:

• Developed a two-stage framework: multi-dimensional feature fusion and attention-guided refinement.
• Designed a multi-stream subnetwork (MFNet) for spatial, angular, EPI, and pseudo-video feature extraction.
• Incorporated a geometry-prior-based subnetwork (GSPNet) and an attention-guided fusion subnetwork (AFNet).

Main Results:

• The LFAMF model significantly outperforms state-of-the-art methods on synthetic and real-world datasets.
• Demonstrated superior performance in maintaining structural integrity at occlusion boundaries and highly textured areas.
• Achieved enhanced angular consistency in reconstructed light fields.

Conclusions:

• The proposed LFAMF network offers a robust solution for light field angular reconstruction.
• The attention-guided fusion mechanism effectively integrates multi-dimensional features for improved results.
• LFAMF advances the state-of-the-art in handling complex scenarios for angular super-resolution.