Real-World Light Field Image Spatial Super-Resolution Based on Multi-View High-Dimensional Geometric Consistency Constraints

In recent years, deep learning-based light field (LF) super-resolution (SR) methods have significantly improved LF image quality. However, most existing LFSR rely on fixed bicubic downsampling to construct SR models, which limits their performance on real-world LF with diverse degradations. To address this issue, we propose an effective SR architecture for real-world LF, termed LRVCSR, which simulates the optimization process to solve real-world LF degradations in a data-driven manner. Specifically, we construct a practical degradation model to characterize the degradation process of real-world LF and solve it by leveraging the high-dimensional geometric consistency of LF. To better capture the high-dimensional nature of LF, we design an implicit feature interaction module, which maps 4D LF to multiple 2D sub-spaces to capture various forms of low-dimensional LF features. Subsequently, LRVCSR leverages LF geometric consistency and non-local spatial similarity to construct a robust low-rank matrix, suppressing random noise during low-rank recovery. Finally, based on 4D spatial-angular consistency, we design a multi-view consistency recovery module with joint 3D convolutions, which simultaneously enhances features and performs spatial SR. Moreover, we design a self-supervised view consistency loss that exploits intrinsic LF geometric consistency as a strong prior. This loss guides the model to achieve multi-view consistent reconstruction without requiring ground truth, making it adaptable to real-world LF with diverse degradations. Experimental results demonstrate that, compared with state-of-the-art LFSR methods, our method achieves superior reconstruction performance on LF with diverse degradations and exhibits better adaptability to real-world unknown LF.