1. Light field spatial super-resolution (LFSSR)

Improve spatial resolution of each sub-aperture image (SAI).

1.1. CNN-based method

Use convolutions to learn spatial / angular correlations and fuse high-frequency details.

• Residual CNNs: Learn directional features and fuse sub-pixel details (e.g., resLF).
• Feature alignment: Optical-flow-based alignment, deformable conv alignment.
• 4D / separable CNNs: Use 4D conv or spatial–angular separable conv to jointly extract features.
• View fusion models: “All-to-One”, multi-view complementary information fusion.
• Attention-based CNNs: Channel/view attention, angular deformable alignment.

resLF

seperable Conv model

“All-to-One” model

view+channel attention model

1.2. Transformer-based method

Use attention to model long-range spatial-angular dependency.

• Spatial–angular transformer: Self-attention along EPI lines to capture parallax geometry.
• Volume and cross-view transformers: Model correlations across many viewpoints.
• Multi-scale angular transformer: Robust to disparity variations.

EPI attention

volume transformer and cross-view transformer

LF-DET

2. Light field angular super-resolution (LFASR)

Increase number of viewpoints (more SAIs) while preserving geometry.

2.1. Depth-dependent method

Estimate depth/disparity → warp existing views → blend new views.

• Optical-flow and superpixel-based warping
• Layered depth representations
• EPI-based geometry modeling
• Depth-guided warping with occlusion reasoning

2.2. Depth-independent method

Avoid explicit depth; rely on signal priors or learning-based angular patterns.

• CNN-based angular detail restoration (on EPIs)
• Angular attention models to reconstruct views

3. Light field spatial-angular super-resolution (LFSASR)

Simultaneously increase both spatial and angular resolution → full 4D reconstruction.

3.1. Deep learning-based method

Jointly model 4D light field structure (geometry + appearance).

• 4D CNN encoder–decoder
• Pseudo-4D convolution combining EPI + spatial-angular blocks
• Disentangled models: separate spatial and angular subspaces
• EPI-based networks (CNN + LSTM) preserving geometry
• Self-supervised or domain-generalized models for wild light fields

3D encoder

EPI-based networks

Experiments