As autonomous landing emerges as a pivotal technology for deep space missions, the capability to capture high-resolution terrain imagery in real time is essential for precise navigation, obstacle detection, and safe touchdown. However, stringent constraints on mass, power, and optical design impose a trade-off for single-camera systems on landers, limiting their ability to simultaneously achieve wide-angle coverage and fine spatial detail. Moreover, complex degradation patterns inherent to deep-space imaging conditions, including optical aberrations, sensor noise under low illumination, and mechanical assembly tolerances, can induce misleading textures during single-image super-resolution, potentially compromising landing accuracy. To address these challenges, this study presents a dual-focus collaborative super-resolution framework that combines the broad scene overview from a short-focus camera with the detailed textures by a long-focus camera, enhancing reconstruction reliability and minimizing spurious textures.

We first constructed the DFSR-Lunar dataset by selecting high-resolution lunar surface strip images captured by the Lunar Reconnaissance Orbiter’s Narrow Angle Cameras (LROC-NACs) between 2012 and 2019. The dataset encompasses diverse terrain types, including impact craters, mare regions, and crater chains. Following radiometric and brightness corrections, each strip image was cropped to 2000 pixel×2000 pixel to serve as ground truth (GT), yielding a total of 450 images. To simulate realistic dual-focus imaging, a three-stage degradation pipeline was implemented. 1) Optical blur modeling: within a validated optical simulation environment, random micrometer-scale decentering and millidegree-scale tilt errors were introduced into both the short-focus and long-focus lens assemblies. Wavelength-dependent point spread functions (PSFs) were computed across F-, d-, and C-spectral lines and across 40 field angles, then convolved with GT images to emulate diffraction and aberration effects. 2) Mixed Poisson-Gaussian noise: to reflect low-light sensor conditions on the lunar surface, a combined noise model applied Poisson noise (modeling photon shot noise and dark current noise) and Gaussian noise (modeling thermal and quantization noise). 3) Dual-focus pair generation: the “degraded” short-focus images were downsampled to simulate the resolution reduction caused by large pixel size, while long-focus views were obtained by centrally cropping 1000 pixel×1000 pixel window with randomized shifts (0?3 pixel) to simulate mechanical misalignments. This process ensured distinct spatial resolutions and fields of view (FoV), producing aligned short-focus and long-focus image pairs. Subsequently, we proposed a dual-focus super-resolution reconstruction model. Our multi-stage progressive network took a low-resolution short-focus image (LR) and a reference long-focus image (Ref) as inputs. The pipeline comprised: 1) a frequency-aware feature modulation block (FAFMB) integrating a Fourier-domain branch for global frequency information and a dual-branch spatial module for non-local/local detail aggregation; 2) a texture similarity module using pre-trained VGG19 (first seven layers) to extract features, compute patch-wise similarity, and generate an index map and confidence map; 3) a deformable convolution alignment block (DCN align) that first warped long-focus features based on the index map, then refined offsets via learned deformable convolutions for precise geometric alignment of Ref features to LR geometry; 4) an adaptive fusion module that learned pixel-wise fusion weights, guided by the confidence map, to merge aligned features, followed by a skip-connection with the original LR features; and 5) a separate single-image super-resolution branch (SISR) processed only the short-focus input through a series of residual channel-attention blocks in parallel, ensuring stable detail enhancement when correspondence confidence is low. The network was trained on 400 image pairs (with rotation and flip augmentations) over 100 epochs using L1 loss and Adam optimizer (initial learning rate of 1×10^-4, halved at scheduled epochs), with input patches of 128 pixel×128 pixel for short-focus views and 256 pixel×256 pixel for long-focus views.

On the held-out test set (50 image pairs), our method outperformed bicubic interpolation, state-of-the-art SISR models (RCAN, SwinIR), and reference-based approaches (TTSR, DCSR) in both PSNR and SSIM metrics (Table 1). Specifically, we achieved an average PSNR of 33.380 dB and SSIM of 0.887, representing improvements of 0.172 dB and 0.012 over the second-best method (DCSR) respectively. Qualitative results of lunar crater rims demonstrated that our framework successfully reconstructed fine rim and rock textures while suppressing noise and spurious edges (Fig. 9). Moreover, modulation transfer function (MTF) analysis revealed significant gains in mid-to-high frequency response across both center and edge FoVs, indicating superior preservation of spatial frequencies compared to degraded inputs and competing methods (Fig. 10). Ablation studies further validated each component’s contribution: removing the FAFMB block (w/o FAFMB) reduced PSNR by 0.338 dB and SSIM by 0.024, while omitting the SISR branch (w/o SISR) yields comparable reductions in perceptual quality (Table 2).

This work introduces a dedicated dual-focus dataset and corresponding reconstruction network for deep-space lander imaging. By realistically modeling optical and sensor degradations and combining complementary camera views through frequency-aware feature modulation, deformable convolution alignment, and adaptive fusion alongside an independent short-focus super-resolution path, the approach delivers more dependable reconstruction of lunar terrain imagery. Future work will extend this framework to diverse planetary environments and optimize computational efficiency for onboard deployment.