LMF-Net: A Lunar Multifocus Image Fusion and High-Fidelity Detail Reconstruction Network for Double-Blur Transition Regions

High-quality lunar imaging is essential to terrain modeling, resource assessment, and landing navigation. However, due to the limited depth-of-field of lunar probe cameras, a single image cannot simultaneously capture clear near-field details and far-field structures. Although multifocus image fusion (MFIF) alleviates this issue, in lunar scenarios, existing methods are hindered by factors, such as extreme illumination conditions, weak textures, and the double-blur phenomenon near focus–defocus boundaries, which degrade focus discrimination and detail reconstruction. To address these challenges, a two-stage lunar MFIF and high-fidelity detail reconstruction network is proposed. Specifically, in the initial fusion stage, multiscale feature extraction combined with an improved spatial-frequency-based decision strategy enables robust focus estimation and precise localization of double-blur transition regions, producing a coarse fused image. In the regional fusion stage, the network reconstructs fine texture details within the transition regions and adaptively integrates them with the coarse fused image, producing visually natural all-in-focus images and enhancing global consistency. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches in both qualitative and quantitative evaluations. Specifically, it achieves a 2.17 dB improvement in peak signal-to-noise ratio and an approximately 22.6% increase in average gradient over the best competing approach, validating its effectiveness and robustness for high-fidelity lunar image fusion under complex illumination conditions.