Directionally factorized light field reconstruction with cross-epipolar and spatial modeling

High-angular-resolution light fields (LFs) are essential for advanced imaging tasks such as post-capture refocusing, depth estimation, and virtual reality. However, acquiring densely sampled LFs remains challenging due to hardware limitations and high capture costs. We propose DFCNet, an end-to-end framework for LF angular super-resolution that reconstructs high-resolution LFs from sparse inputs. The network leverages a hierarchical Residual-in-Residual architecture with Directionally Factorized Convolutional Blocks to disentangle spatial, epipolar, and cross-epipolar features while preserving angular consistency. Channel recalibration via Squeeze-and-Excitation blocks adaptively emphasizes informative features, and a hybrid interpolation-reconstruction strategy ensures accurate and stable angular upsampling. Extensive experiments on synthetic and real-world datasets demonstrate that DFCNet consistently outperforms state-of-the-art methods in average PSNR and SSIM, accurately reconstructing fine textures, occluded regions, and angular structures. The method also generalizes effectively to extrapolation tasks, synthesizing previously unseen views with high fidelity. These results establish DFCNet as a robust and efficient solution for high-quality light field reconstruction.