A novel EPI-guided network with progressive fusion for light field reconstruction

Dense light fields (LFs) hold significant potential in computer vision. However, since the existing LF acquisition devices struggle to capture high-quality dense LFs, researchers have proposed computational methods to reconstruct dense LFs from sparse LFs. Nevertheless, existing methods usually learn multiple features of the LF equally to achieve dense LF reconstruction while overlooking the discrepancies and correlations between different features. The specificity of epipolar plane images (EPIs) features as an important bridge between the LF spatial and angular information is worth exploring in depth. In this paper, we propose a novel EPI-guided network to emphasize differentiated learning of EPI features. The network extracts new guiding information from EPI global features and guides spatial and angular features to establish spatial-angular correlations of LF data effectively. We also introduce a progressive feature fusion strategy, which sequentially fuses the LF spatial, angular, and EPI features. This strategy fully explores the correlations between each pair of features and achieves differentiated learning among different features. The quantitative and qualitative experimental results demonstrate that our network achieves state-of-the-art reconstruction performance on large and small disparities datasets. Codes will be released at https://github.com/Baoshuai129/EGPFNet.