Iterative Token Refinement for Remote Sensing Image Super-Resolution
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Recent advances in remote sensing image super-resolution (RSISR) have leveraged transformer architectures to achieve state-of-the-art results. Nevertheless, most existing transformer-based approaches treat the token sequence as a unified global context, which often fail to preserve critical local structures and cross-region semantics. To address these issues, we propose an iterative transformer framework that progressively refines visual tokens with contextual awareness. Specifically, we initially represent an input image as discrete tokens based on vector quantized (VQ)-GAN. Then, building upon this tokenized representation, we propose a predictor–evaluator mechanism that iteratively refines the high-resolution (HR) token sequence. In each iteration, the predictor generates all HR tokens and retains those with the highest confidence estimates, while masking the remainder for subsequent refinement. Concurrently, the evaluator, which is trained to assess the reliability of each token, guides the selection and regeneration of tokens to enable progressive reconstruction. To preserve structural consistency with the low-resolution (LR) input, we further introduce a Consistency-aware LR-conditioned Decoder module, which leverages the LR image to regulate the HR tokens. Experiments show that our method achieves competitive performance on multiple remote sensing benchmarks.
Fields of Study & Subjects
Advanced Image Processing Techniques
Remote-Sensing Image Classification
Generative Adversarial Networks and Image Synthesis
Publication details
- Journal / source
- IEEE Geoscience and Remote Sensing Letters
- Volume
- 23
- Pages
- 6007105–6007105
- Publication date
- 1 January 2026
- Language
- EN