History-Augmented Contrastive Meta-Learning for Unsupervised Blind Super-Resolution of Planetary Remote Sensing Images PubID: pubid: 0000–0000/00$00.00 © 2021 IEEE
Huijia Zhao, Student Member, IEEE, Jie Lu, Yunqing Jiang, Xiao-Ping Lu*, Kaichang Di · 2025
Read History-Augmented Contrastive Meta-Learning for Unsupervised Blind Super-Resolution of Planetary Remote Sensing Images PubID: pubid: 0000–0000/00$00.00 © 2021 IEEE by Huijia Zhao, Student Member, IEEE, Jie Lu, Yunqing Jiang, Xiao-Ping Lu*, Kaichang Di (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-20045.txt.
Introduction
Hui-Jia Zhao, Jie Lu, Yunqing Jiang, Xiao-Ping Lu, and Kaichang Di present HAC-MoE, an unsupervised method for sharpening planetary images. Blind super-resolution of remote sensing pictures is ill-posed: the blur is unknown and no ground-truth images exist for supervision. Older unsupervised methods suffer unstable training and shifting distributions, and they blur distinct surface features.
The framework separates blur kernel estimation from image reconstruction with no outside kernel priors. A contrastive sampling step builds plausible kernels, and a history-augmented contrastive strategy uses past model states as negative self-priors, which the authors show steadies training. A morphology-aware mixture of experts then rebuilds varied planetary terrain. Their new Ceres-50 benchmark supports the tests, where the method leads in reconstruction and kernel accuracy.
Catalog record
| CREATOR | Huijia Zhao, Student Member, IEEE, Jie Lu, Yunqing Jiang, Xiao-Ping Lu*, Kaichang Di |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-20045
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
History-Augmented Contrastive Meta-Learning for Unsupervised Blind Super-Resolution of Planetary Remote Sensing Images PubID: pubid: 0000–0000/00$00.00 © 2021 IEEE Huijia Zhao, Student Member, IEEE, Jie Lu, Yunqing Jiang, Xiao-Ping Lu*, Kaichang Di Abstract Planetary remote sensing images are affected by diverse and unknown degradations caused by imaging environments and hardware constraints. These factors limit image quality and hinder supervised blind super-resolution due to the lack of ground-truth images. This work presents History-Augmented Contrastive Blind Super-Resolution (HACBSR), an unsupervised framework for blind super-resolution that operates without ground-truth images and external kernel priors. HACBSR comprises two components: (1) a contrastive kernel sampling mechanism with kernel similarity control to mitigate distribution bias from Gaussian sampling, and (2) a history-augmented contrastive learning that uses historical models to generate negative samples to enable less greedy optimization and to induce strong convexity without ground-truth. A convergence analysis of the history-augmented contrastive learning is given in the Appendix. To support evaluation in planetary applications, we introduce Ceres-50, a dataset with diverse geological features simulated degradation patterns. Experiments show that HACBSR achieves competitive performance compared with state-of-the-art unsupervised methods across multiple upscaling factors. The code is available at https://github.com/2333repeat/HACBSR , and the dataset is available at https://github.com/2333repeat/Ceres-50 . I Introduction Planetary remote sensing images serve as a foundation for understanding the geological and morphological characteristics of celestial bodies beyond Earth. These images enable scientists to conduct comparative planetary studies, analyze surface features, and investigate the geological evolution of planets and moons throughout the solar system. However, acquiring high-quality planetary images presents technical challenges that differ from terrestrial remote sensing. Deep space missions operate under severe constraints that directly impact imaging capabilities. Spacecraft payload limitations restrict the size and complexity of imaging systems, while extreme distances introduce significant communication delays and bandwidth restrictions. Power constraints further limit the operational capacity of imaging sensors, forcing trade-offs between spatial resolution, temporal coverage, and data transmission rates. These engineering realities result in planetary images that often exhibit insufficient resolution for detailed scientific analysis. The imaging environments across different planetary bodies present unique degradation patterns that compound these technical limitations. Martian observations contend with frequent dust storms that introduce atmospheric scattering and obscure surface features. Venus presents an even more challenging scenario, where dense sulfuric acid clouds and extreme surface temperatures create complex optical distortions. The icy moons of Jupiter and Saturn introduce additional complications through ice crystal scattering, plasma interference from radiation belts, and tidal deformation effects. Each planetary environment requires adaptive image enhancement strategies that can accommodate diverse and often unknown degradation processes. To address these requirements, super-resolution (SR) has emerged as a viable solution. However, early SR methods primarily focused on bicubic degradation [ 4 , 10 , 16 , 32 , 3 ] . Our prior work also investigated neural operators in planetary remote sensing SR under bicubic downsampling [ 46 ] . However, degradation in real planetary remote sensing images is unknown and complex. Furthermore, blind super-resolution (BSR) represents a more realistic framework for planetary image enhancement, as it addresses scenarios where the degradation process is unknown or partially observable. A typical degradation process of BSR can be described as follows: \mathbf{y}=(\mathbf{x}\otimes\mathbf{k})\downarrow_{s}+\mathbf{z}, (1) where \mathbf{y}\in\mathbb{R}^{H\times W} is the observed low-resolution (LR) image, \mathbf{x}\in\mathbb{R}^{sH\times sW} is the unknown high-resolution (HR) image, \otimes denotes 2D convolution, \downarrow_{s} denotes downsampling by factor s , \mathbf{k}\in\mathbb{R}^{K\times K} is an unknown blur kernel, and \mathbf{z} denotes the additive noise. Beyond the degradation, current BSR methods can be broadly grouped into two paradigms: joint optimization approaches that simultaneously estimate degradation kernels and reconstruct high-resolution images [ 1 , 10 , 41 , 9 , 14 ] , and alternating optimization frameworks that decouple kernel estimation from image reconstruction [ 13 , 28 , 35 , 37 ] . While these methods report strong results on natural images, their application to planetary remote sensing remains limited. Fig. 1: Comparison of kernel distribution between random kernel samples and contrastive kernel sampling. The contrastive kernel sampling effectively generates kernels with balanced distribution. It also reduces the probability of generating outlier kernels, which can adversely affect image reconstruction. First, existing BSR methods rely heavily on supervised training. Most deep learning-based BSR methods are constrained by their dependence on ground-truth images or kernel priors [ 17 , 44 , 39 , 42 , 41 ] . However, in planetary remote sensing, acquiring large datasets is challenging and ground truth images are impossible to obtain. Therefore, developing a method that does not rely on supervised training is essential. Second, Planetary remote sensing images exhibit diverse and often unknown degradations due to dramatically different imaging environments across celestial bodies. Mars observations are compromised by dust storms that create atmospheric scattering and infrared thermal radiation artifacts, while Venus imaging must contend with dense cloud cover occlusion, sulfuric acid haze, and extreme high-temperature effects. Jupiter’s moons present different degradation patterns including ice crystal scattering, plasma interference, and tidal deformation distortions. Different planetary remote sensing images exhibit different degradation patterns. Therefore, it is necessary to develop a BSR method that does not rely on presupposed kernel priors. Third, the planetary remote sensing domain lacks BSR datasets. The lack of planetary remote sensing datasets significantly hinders development in this area. The evaluation of planetary remote sensing image restoration tasks also requires BSR datasets. Although existing BSR methods have achieved strong performance, early research has focused primarily on reconstructing HR images and estimating image degradation with a single model [ 44 , 39 , 42 , 41 , 2 ] . Separating the image degradation estimation and HR image reconstruction processes has emerged as an alternative approach and has also achieved strong performance [ 1 , 5 , 28 ] . However, most BSR methods are trained and evaluated on datasets sharing the same degradation kernel distribution, often assuming that training and test kernels are matched or relying on kernel priors [ 27 , 44 , 10 ] . When these methods encounter training and test datasets with different degradation models, their performance degrades. Therefore, an unsupervised BSR method is needed that can reconstruct images without presupposed degradation kernels and handle images with different degradation kernels. Existing unsupervised BSR methods generate kernels primarily by sampling key parameters from uniform distributions [ 35 , 37 , 13 ] , which may lead to unbalanced kernel distributions, as shown in Figure 1 , potentially resulting in outlier kernels that are detrimental to reconstruction. Furthermore, since ground truth images do not participate in training and inference, the lack of supervised information may cause overly aggressive optimization, potentially leading to overfitting. To overcome these limitations, we propose History-Augmented Contrastive Blind Super-Resolution (HACBSR), an unsupervised BSR model that does not rely on ground-truth images or degradation kernel priors. HACBSR incorporates contrastive kernel sampling in the kernel learning (KL) stage and history-augmented contrastive learning with meta-learning in the contrastive image learning (CIL) stage. Unlike previous alternating optimization unsupervised BSR approaches, the contrastive kernel sampling overcomes the kernel bias generated from random explicit Gaussian sampling, as shown in Fig. 1 , and balances the kernel distribution to provide improved kernel priors and achieve more general parameters for the next stage of optimization. The history-augmented contrastive learning provides self-prior guidance from the historical image estimator