Image Super-Resolution
Overview
Image super-resolution estimates sharper, higher-resolution images from one or more low-resolution observations, supporting photography, microscopy, remote sensing, and medical imaging when direct high-resolution capture is limited. Methods range from CNNs and attention or state-space Transformers to continuous implicit representations and multi-frame fusion; diffusion, flow-matching, and autoregressive models generate plausible fine detail, with distillation and quantization helping reduce inference costs. Evaluation spans synthetic benchmarks using PSNR and SSIM, real-world perceptual measures and human judgments, and task-specific tests such as recognition or segmentation. A central tension is that realistic textures can be hallucinated or depart from observed content, while synthetic training and pixel metrics may not predict performance on real sensors.
Latest papers
249 papers23 papers in the last four weeks, up 5% on the four weeks before. 0.1% of all new papers.