Super-Resolution

Super-resolution reconstructs a sharp high-resolution image from a small blurry one by learning what fine detail typically hides behind coarse pixels.

Super-resolution learns plausible high frequencies from low-resolution input, trading pixel fidelity against perceptual sharpness

Why Does This Exist?

Zooming is not resolving. Bicubic interpolation stretches $128$ pixels into $512$ smoothly, inventing no eyelashes, no brick lines, no texture, just larger blur. Surveillance, medical imaging, and photo restoration need actual detail, plausible high frequencies consistent with the low-resolution evidence.

Super-resolution learns the prior over details from data: given this blurry patch, sharp patches in the training set looked like these. The task family map is image-to-image tasks, the founding convolutional method is SRCNN, and the perceptual-quality champion is ESRGAN.

Think of It Like This

A restorer filling cracks in a fresco

A fresco restorer sees a damaged angel wing: coarse shape survives, feather barbs are gone. They study intact wings across the chapel and repaint barbs that fit this wing's pose and light, barbs the original surely had, though no record proves each stroke.

The low-resolution image is the damaged wing and the training set is the chapel. Bicubic interpolation is plastering the crack flat. The restorer fails when chapels differ, training priors from faces smeared onto text, which is the domain limit below.

How It Actually Works

The forward model downsamples: $y = DHx + n$ , with blur $H$ , decimation $D$ , and noise $n$ . Single-image super-resolution inverts it with learned priors at scale factor $s$ , usually $2\times$ or $4\times$ . Two philosophies compete: fidelity methods minimise pixel error (PSNR-oriented), perceptual methods add adversarial and feature losses for sharpness humans prefer despite worse pixels.

Standard scoring uses PSNR $= 10 \log_{10}(\mathrm{MAX}^2 / \mathrm{MSE})$ and SSIM for structure. The perception-distortion tradeoff is fundamental: the sharpest-looking outputs rarely minimise MSE, because MSE rewards safe averages over bold correct textures.

Worked example

Upscale $128 \to 512$ ( $4\times$ ). A reconstruction with MSE $0.004$ on $[0, 1]$ pixels gives PSNR $10 \log_{10}(1/0.004) = 10 \log_{10}(250) = 23.98$ dB. A blurrier output at MSE $0.002$ scores $26.99$ dB and looks worse, while a crisp GAN output at MSE $0.006$ scores $22.22$ dB and looks better. The ranking inverts because MSE cannot see sharpness, only averages.

Code

PSNR arithmetic that prices the fidelity-versus-sharpness tradeoff.

1import numpy as np2
3
4def psnr(mse: float, maxval: float = 1.0) -> float:5    """Peak signal to noise ratio from mean squared error."""6    return 10 * float(np.log10(maxval ** 2 / mse))7
8
9print(round(psnr(0.004), 2))  # -> 23.98 dB our sharp example10print(round(psnr(0.002), 2))  # -> 26.99 dB blurrier but higher11print(round(psnr(0.006), 2))  # -> 22.22 dB crisp GAN output, lowest

Watch Out For

Chasing PSNR into blur

Leaderboards ranked by PSNR alone select the blurriest competitive model, since averaging candidates minimises squared error. Teams ship record numbers that users call waxy.

Report perceptual metrics (LPIPS, DISTS) and human ratings beside PSNR. State which tradeoff point the model targets instead of claiming both crowns.

Training on bicubic downsamples, shipping on real blur

Synthetic $y$ from bicubic decimation mismatches camera blur, sensor noise, and compression. Models excel on benchmarks and smear real photos, the classic domain gap.

Train with realistic degradation pipelines or real paired captures. The degradation model matters more than the network past a point.

The Quick Version

Super-resolution inverts blur plus decimation using learned detail priors.
Our $4\times$ example at MSE $0.004$ gives PSNR $23.98$ dB.
Blurrier outputs can outscore crisp ones: $26.99$ versus $22.22$ dB.
Fidelity and perceptual quality trade off fundamentally; report both.
Bicubic-trained models gap on real degradations; model the pipeline realistically.