S2-Net: Synergizing Spectral Compensation and Spatial Structure for Light Field Reconstruction

Abstract

Due to limitations in optical design and acquisition cost, densely sampled light fields (LFs) cannot be directly captured in practice, motivating LF angular reconstruction from sparsely sampled views. Reconstructing densely-sampled LFs from sparse views necessitates simultaneously modeling long-range epipolar geometry and preserving fine-grained textures. However, we identify a fundamental spectral bias in existing State-of-the-Art (SOTA) Transformer-based approaches: while global attention mechanisms excel at capturing large-disparity dependencies, they inherently act as low-pass filters, causing spectral collapse in high-frequency regions (e.g., textures and edges). To bridge this gap between global geometric consistency and local spectral fidelity, we propose a novel dual-domain calibration framework termed S2-Net, which synergizes Spectral compensation and Spatial structure. Specifically, we introduce a Spectral Residual Modulation (SRM) module that constructs an adaptive band-pass filter in the frequency domain and reverses the spectral attenuation caused by global attention via a learnable spectral mask. Furthermore, we introduce a Structural Prior Guidance (SPG) module in the spatial domain to inject an explicit structural prior, where a gating mechanism adaptively regulates feature propagation, suppressing texture over-smoothing while preserving geometric consistency in the feature space. Extensive experiments on synthetic and real-world LF datasets demonstrate that the proposed method consistently outperforms SOTA approaches in both reconstruction accuracy and visual quality.