FourASR: Exploiting Fourier Frequency Information for Efficient Light Field Angular Super-Resolution

Light field (LF) angular super-resolution (SR) aims to reconstruct a densely sampled LF from a sparse input. A key challenge is the large disparity between sparse views, which disrupts angular consistency and makes capturing cross-view correlations difficult. Disparity-based methods often yield errors in textureless, occluded, or reflective regions, causing ghosting and edge misalignment. A well-established solution should disentangle the 4D LF into three 2D subspaces-the subaperture image (SAI), horizontal epipolar plane image (HEPI), and vertical epipolar plane image (VEPI). While non-disparity-based methods avoid explicit disparity estimation, they typically process these subspaces using local convolutions with limited receptive fields. To address this, we propose an Adaptive Frequency Selection Neural Operator (AFSNO) to exploit global spatial relationships via frequency-domain modeling. Inspired by the convolution theorem, AFSNO performs adaptive frequency selection through a learnable mask, acting as a global convolution in the spatial domain. However, directly applying AFSNO to numerous LF subspaces incurs significant parameter overhead and lacks resolution flexibility. To overcome this, we introduce a 2D Subspace Patching Strategy (2DSPS) that enables weight sharing across patches. Specifically, 2DSPS tailors the patching scheme to the distinct characteristics of the SAI, HEPI, and VEPI subspaces, instantiating specialized AFSNO operators for each. This strategy drastically reduces model complexity while allowing the network to handle varying input resolutions during inference. Combined with local spatial convolutions, these components form the proposed Four-Block for efficient LF feature enhancement, constituting a simple yet powerful framework called FourASR. Additionally, to better reconstruct high-frequency details, we introduce a Fourier frequency-domain loss. Extensive experiments demonstrate that FourASR outperforms state-of-the-art methods significantly in both quantitative metrics and perceptual quality.