Spatial-Angular Decoupling Interaction Networks for Light Field Angular Super-Resolution

The light field image simultaneously captures both the intensity and directional information of light rays, making it a valuable tool for acquiring four-dimensional spatial data. However, the restriction of sensor resolution results in a trade-off between angular and spatial resolution, which hinders our ability to simultaneously acquire light field with high spatial and angular resolution. In this paper, we focus on the sparse light field reconstruction task, which aims to construct densely-sampled light fields for higher angular resolution. To achieve this goal, we propose a spatial-angular decoupling interaction network. The method is based on the light field decoupling mechanism, which interactively fuses and updates spatial and angular features to achieve global spatial and angular features. Moreover, four directions epipolar plane features are extracted to obtain more disparity information. Finally, the global spatial and angular features are interacted with the fused epipolar plane features, and the dense views light field is reconstructed by the fused features. Experimental results indicate that the spatial-angular decoupling interaction network outperforms other state-of-the-art methods in angular domain reconstruction, and achieves more refined texture details in the reconstructed views.