Texture-Enhanced Light Field Super-Resolution With Spatio-Angular Decomposition Kernels

Despite the recent progress in light field super-resolution (LFSR) achieved by convolutional neural networks, correlation information of LF images has not been sufficiently studied and exploited due to complexity 4-D data. To cope with such high-dimensional data, most existing LFSR methods resorted decomposing it into lower dimensions subsequently performing optimization on decomposed subspaces. However, these are inherently limited as they neglected characteristics decomposition operations only utilized a set subspaces ending up failing extract spatio-angular features leading performance bottleneck. overcome limitations, this article, we comprehensively discover potentials propose novel concept kernels. In particular, systematically unify various series kernels, which incorporated our proposed kernel network (DKNet) for comprehensive feature extraction. The DKNet is experimentally verified achieve considerable improvements compared state-of-the-art methods. further improve producing more visually pleasing results, based VGG network, (LFVGG) loss guide texture-enhanced (TE-DKNet) generate rich authentic textures enhance images’ visual quality significantly. We also an indirect evaluation metric taking advantage material recognition objectively assess perceptual enhancement brought LFVGG loss.