Super-resolution of light field images based on hierarchical interactive learning

To address the inherent spatial-angular resolution trade-off in light field image super-resolution, this paper proposes a novel Hierarchical Interactive Learning Network (HILN). The network synergistically integrates multi-scale feature extraction with hierarchical interactive learning. It employs an epipolar plane attention module to capture long-ranges patial-angular dependencies and introduces a Hierarchical Dynamic Channel Attention (HDCA) mechanism. The HDCA utilizes an efficient channel attention module at shallow layers to enhance high-frequency details through local channel interaction and dynamic kernel adjustment. At deeper layers, it incorporates a dynamic channel attention module that employs parameterized weight prediction for adaptive integration of cross-view features, significantly improving view consistency. Extensive experiments on five standard benchmarks demonstrate that our method outperforms state-of-the-art approaches in both 2× and 4× super-resolution tasks in terms of PSNR and SSIM. Notably, on the challenging STF gantry scene, our method achieves a PSNR of 32.26 dB for 4× super-resolution. Ablation studies further confirm the effectiveness of each component, showing that removing shallow or deep modules leads to differentiated performance impacts across scenarios, which validates the rationality of the hierarchical design and the complementary nature of the modules.