Light Field Angular Super-Resolution Network Based on Convolutional Transformer and Deep Deblurring

Abstract

Many Light Field (LF) angular super-resolution methods have been proposed to cope with the LF spatial and angular resolution trade-off problem. However, most existing methods cannot simultaneously explore LF local and non-local geometric information, which limits their performances. Moreover, since the quality degradation model of the reconstructed dense LF is always neglected, most solutions fail to effectively suppress the blurry edges and artifacts. To overcome these limitations, this paper proposes an LF angular super-resolution network based on convolutional Transformer and deep deblurring. The proposed method mainly comprises a Global-Local coupled Convolutional Transformer Network (GLCTNet), a Deep Deblurring Network (DDNet), and a Texture-aware feature Fusion Network (TFNet). The GLCTNet can fully capture the long-range dependencies while strengthening the locality of each view. The DDNet is utilized to construct the quality degradation model of the reconstructed dense LF to suppress the introduced blurred edges and artifacts. The TFNet distills the texture features by extracting the local binary pattern map and gradient map, and allows a sufficient interaction of the obtained non-local geometric information, local structural information, and texture information for LF angular super-resolution. Comprehensive experiments demonstrate the superiority of our proposed method in various LF angular super-resolution tasks. The depth estimation application further verifies the effectiveness of our method in generating high-quality dense LF.