Self-Supervised Learning for Spatial-Domain Light-Field Super-Resolution Imaging

This paper proposes a selfsupervised learningbased method for the superresolution imaging of spatialdomain resolutionlimited lightfield images. Using deep learning selfencoding, a superresolution reconstruction of the spatial-domain is performed simultaneously for all light field subaperture images. A hybrid loss function based on multiscale feature structure and total variation regularization is designed to constrain the similarity of the model output image to the original lowresolution image. Numerical experiments show that the newly proposed method has a suppressive effect on noise, and the resultant average superresolutions for different light field imaging datasets exceed those of the supervised learningbased method for light field spatial domain images.