Pyramid Mamba with Multi-directional Scanning for Light Field Image Super-Resolution

Light field imaging has emerged as a revolutionary technology with widespread applications in fields such as virtual reality and robotics. However, enhancing the resolution of light field images presents several challenges. To address these issues, this study proposes a novel illumination-aware Light Field Image Super-Resolution framework that effectively leverages a hybrid encoder-enhancer-decoder architecture. Our framework employs a multi-stage processing pipeline, which includes pre-enhancement through Retinex-based illumination normalization, feature extraction via a shallow convolutional head, and angular modeling using an Enhanced Angle Mamba (EAM) module that captures angular dependencies across multiple scales. Furthermore, we introduce a spatial refinement stage utilizing RRDB-lite blocks to recover high-frequency details, accompanied by a spatial-channel fusion module that synergistically merges angular and spatial features. Our method effectively tackles the challenge of angular feature extraction in light field images. Experimental results conducted on seven publicly available datasets demonstrate that our LFISR framework significantly outperforms existing state-of-the-art methods based on CNNs, Transformers, and Mamba, achieving superior visual quality and robustness, as demonstrated by enhancements in PSNR and SSIM. These findings validate the efficacy of our model and highlight its potential for enhancing light field image quality in practical applications.