View position prior-supervised light field angular super-resolution network with asymmetric feature extraction and spatial-angular interaction

Light field imaging can record the intensity and direction information of light rays in space, which attracts extensive attention. However, the trade-off between angular and spatial resolution cannot be avoided due to the limitation of sensors in commercial light field cameras. To mitigate this problem, this paper proposes the view position prior-supervised light field angular super-resolution network with asymmetric feature extraction and spatial-angular interaction. First, there is a severe information asymmetry between the spatial and angular dimensions in light fields with sparse views. The asymmetric feature extraction block is proposed to extract spatial and angular features with different receptive fields in an asymmetric manner. As a result, more light field intrinsic features are extracted, which improves the utilization rate of the limited light field information. Second, the existing methods usually ignore the correlations among the newly synthesized views. The spatial-angular interaction module is proposed to collect the local and global information, build relations between any two points in the feature space, and reconstruct light field consistencies. Thus, the complete view correlations can be established. Last but not least, we investigate the impact of the given views on each viewpoint and propose a loss function based on the view position prior, which reduces the quality difference among the synthesized sub-aperture images and further improves the network performance. Comprehensive experiments demonstrate that our method can perform best on all datasets, and the depth estimation results present a new perspective to show the superiority of the proposed method.