superresolution

Superresolution refers to techniques for reconstructing a higher-resolution image or sequence from one or more lower-resolution observations, with the goal of recovering details beyond the native resolution of the imaging system. In practice, SR seeks to invert the degradation that occurs during image formation, which typically includes blur, subsampling and noise.

A common degradation model expresses a low-resolution image y as y = D H x + n, where x is the high-resolution image, H is a blur operator, D represents downsampling, and n denotes noise. When multiple observations are available, alignments (often sub-pixel shifts) are used to fuse information to recover x more accurately.

Methods are categorized as interpolation-based (simple upsampling; examples include nearest-neighbor, bilinear, bicubic) and, more importantly, reconstruction-based and learning-based approaches. Reconstruction-based SR formulates x as the solution to an inverse problem with regularization or priors. Example-based and exemplar-based methods exploit patches from external data. Learning-based SR uses data-driven models, often trained to map low- to high-resolution patches.

Recent advances have emphasized deep learning. Convolutional neural networks such as SRCNN pioneered end-to-end SR, followed by deeper models (EDSR, RCAN) and generative approaches (SRGAN, ESRGAN) that can generate sharper textures. Multi-frame or video SR uses multiple frames to improve accuracy. Some methods also address blind SR, where the blur kernel is unknown.

Evaluation uses metrics like PSNR and SSIM on benchmarks. Popular datasets for development include Set5, Set14, BSD100 and Urban100. Applications span satellite and aerial imaging, medical imaging, surveillance and consumer photography. Challenges include the ill-posed nature of SR, sensitivity to noise, registration errors, computational demands, and the risk of introducing artifacts or hallucinated details in critical tasks.