2026· IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing· Vol 19, pp. 24520-24533· 0 citations· 55 references
TL;DR
This work proposes a multistep tunable SR network named MTSR, which simulates the mapping process from low-resolution inputs to high-resolution outputs and introduces adjustable parameters, and enables the method to produce nonsmooth SR results without artifacts while preserving rich and realistic textural details.
Abstract
Recently, with the rapid development of deep learning technology, remote sensing image super-resolution (SR) reconstruction has become an important research direction in the field of computer vision. Traditional deep learning-based SR methods often suffer from oversmoothed reconstruction results and lack of high-frequency details. Subsequent methods have been improved to produce SR results that better align with human visual perception. However, these methods tend to introduce unpleasant artifacts into the reconstructed images. Currently diffusion-based approaches leverage forward noise addition and backward denoising processes to generate artifact-free SR images that also conform to visual perception. Nevertheless diffusion-based models still suffer from unstable SR performance. To address these issues, we propose a multistep tunable SR network named MTSR. Our network simulates the mapping process from low-resolution inputs to high-resolution outputs and introduces adjustable parameters. This design enables our method to produce nonsmooth SR results without artifacts while preserving rich and realistic textural details. Meanwhile, this study also optimizes the training network, enabling it to better capture the intermediate information transformations from low-resolution to high-resolution images. We conduct experiments on three datasets, and the experimental results demonstrate that our method achieves excellent texture recovery performance.
A novel SR framework based on the flow matching paradigm and a diffusion transformer, named FlowT-SR, which achieves superior and reliable reconstruction quality by jointly mitigating sensor noise and thin cloud interference, achieving superior reconstruction performance compared with current state-of-the-art methods in terms of both PSNR and SSIM.
Yu-Tong Zhang, Guang Yang, Rong Liu et al.· Italian National Conference...· 0 citations
Abstract. Remote sensing image super-resolution aims to recover fine structural and textural details from degraded low-resolution observations. However, conventional methods and early deep learning models often produce over-smoothed results and struggle to reconstruct realistic high-frequency content. Stable Diffusion-based (SD-based) methods offer a promising alternative by using strong generative priors to synthesize more natural, detail-rich super-resolved images. Although many SD-based super-resolution methods have been proposed in computer vision, their use in remote sensing imagery remains limited, and systematic comparative evaluation in this domain is still lacking, leaving insufficient empirical guidance for method development. Therefore, this paper compares four representative SD-based super-resolution methods, namely Stable Super-Resolution (StableSR), Semantics- Aware Super-Resolution (SeeSR), Different Blind Image Restoration (DiffBIR), and Pixel-Aware Stable Diffusion (PASD), on the WHU-Mix remote sensing dataset. The evaluation uses seven metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), Frechet Inception Distance (FID), CLIP Image Quality Assessment (CLIP-IQA), Multi-Scale Image Quality Transformer (MUSIQ), and Multi-Dimension Attention Network for No-Reference Image Quality Assessment (MANIQA). Quantitative results show that StableSR achieves the highest PSNR of 23.16 dB, PASD obtains the best SSIM of 0.81 and lowest LPIPS of 0.45, SeeSR achieves the best MUSIQ of 64.57 and MANIQA of 0.46, and DiffBIR achieves the best FID of 110.58 and CLIP-IQA of 0.68 but with weaker full-reference fidelity. These findings indicate that current SD-based methods favor different aspects, including fidelity preservation, perceptual quality, and generative realism, and should be selected according to the target remote sensing application.
Yiran Lu, Hongjie He, Fei Zhu et al.· The International Archives o...· 0 citations
Investigation of the use of edge loss to enhance the structural fidelity of SR images for remote sensing imagery indicates that edge loss is an effective and easily implementable strategy for improving SR reconstruction quality in remote sensing imagery.
Minkyung Chung, Youkyung Han· The International Archives o...· 0 citations
Image super-resolution, which aims to reconstruct high-resolution images from their low-resolution observations, is fundamental to medical imaging, remote sensing, surveillance, microscopy, and scientific visualization. Traditional model-based methods formulate super-resolution as an inverse problem with hand-crafted regularization priors. While interpretable and theoretically grounded, they rely on fixed assumptions and require computationally intensive iterative solvers. Deep learning methods offer data-driven flexibility by learning nonlinear mappings from low- to high-resolution images, among which diffusion models have achieved particularly impressive perceptual quality. However, the standard diffusion training objective is a pixel-domain noise-prediction loss that does not explicitly enforce perceptual fidelity, which can lead to oversmoothing and loss of fine image structure. To address these limitations, we propose a perceptually regularized diffusion framework that incorporates prior knowledge through perceptual-loss-based regularization, improving training convergence and encouraging the recovery of meaningful image features. Experiments on benchmark datasets demonstrate improved perceptual quality and competitive distortion metrics, highlighting the effectiveness of regularization for diffusion-based super resolution.
Chuxiang Wang, Pavithra Venkatachalapathy, Ying Liang et al.· 0 citations
Remote sensing image super-resolution is a critical task for reconstructing high-fidelity images from low-resolution observations. However, practical remote sensing scenarios often involve complex, compound degradations that severely compromise essential high-frequency structures, such as sharp edges and fine textures. While conventional convolutional neural network and Transformer-based methods have shown promise, they predominantly rely on static feature representations that lack adaptability to diverse scenes and often struggle to restore specific high-frequency details due to spectral bias. To this end, we propose a novel frequency-aware prompt learning framework. Specifically, a frequency-aware transformer module is designed to leverage the fast Fourier transform for explicit high-frequency component restoration, thereby overcoming the frequency bias inherent in spatial-domain models and effectively recovering intricate structural details. Concurrently, a dynamically generated prompt modulator is introduced to provide scene-specific adaptability via learnable vectors. This allows the network to adaptively calibrate feature responses to mitigate diverse environmental variations and compound artifacts. This synergistic integration ensures superior reconstruction fidelity and robust generalization across heterogeneous remote sensing scenarios. Comprehensive experiments across three benchmark datasets demonstrate that our method achieves outstanding performance in both quantitative metrics and visual quality assessments.
Kangli Zeng, Hang Hu, Ying Yu et al.· IEEE Journal of Selected Top...· 0 citations
Image super-resolution is an important image processing task that aims to reconstruct a high resolution image from a low resolution input. Conventional interpolation methods often produce blurred edges, loss of texture, and pixelated outputs, particularly when images contain noise, compression artefacts, or complex real world degradations. This work proposes a hybrid CNN Transformer GAN framework for natural image super resolution. The CNN component extracts local features such as edges, patterns, and fine textures; the Transformer component captures long range dependencies and preserves global image structure; and the Generative Adversarial Network (GAN) component improves perceptual quality by generating sharper and more realistic visual details. The proposed model supports 2X, 4X, and 8X upscaling, with 4X selected as the primary evaluation scale because it provides a practical balance between image quality and computational complexity. The system is trained using degraded low-resolution images produced through blur, noise, JPEG compression, and downsampling operations. Performance is evaluated using Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), Mean Squared Error (MSE), and visual quality analysis. The hybrid architecture is expected to improve edge sharpness, preserve structural details, reduce visual artefacts, and generate natural looking high-resolution outputs compared with conventional CNN and GAN based approaches. The proposed framework can support applications in photo restoration, mobile imaging, surveillance, e-commerce, digital archives, and web image enhancement.
S. B· International Journal of Lat...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
