This paper presents an improvement on the Mamba-based method, which boasts long-range information modeling and linear complexity, to optimize computational cost and performance in Lightweight Image Super-Resolution (LFSR). To address the inefficient and redundant feature extraction problems of conventional multidirectional scanning strategies applied to complex LF data, this paper designs a Subspace Simple Mamba Block (SSMB) based on the Subspace Simple Scanning (Sub-SS) strategy, achieving more efficient and accurate feature extraction. Furthermore, to address the limitations of state space in preserving spatial-angle and disparity information, a two-stage modeling strategy is proposed to more comprehensively explore nonlocal spatial-angle correlations. In the first stage, the Spatial-Angular Residual Subspace Mamba Block (SA-RSMB) is used to extract shallow spatial-angle features. In the second stage, a dual-branch parallel architecture combining the Epipolar Plane Mamba Block (EPMB) and the Epipolar Plane Transformer Block (EPTB) is used to enhance deep epipolar features. Based on these modules and strategies, we propose LFMT, a hybrid Mamba-Transformer framework that integrates the strengths of the Mamba and Transformer models. LFMT enables comprehensive information exploration across spatial, angular, and epipolar domains. Experimental results demonstrate that LFMT significantly outperforms existing state-of-the-art LFSR methods while maintaining low computational complexity on real and synthetic LF datasets.
