Image Denoising

Image denoising separates true signal from random sensor noise, recovering clean photos by learning what noise looks like at every brightness level.

A residual denoiser predicts the noise layer and subtracts it, lifting PSNR from 20.17 to 28.59 dB while MSE falls from 625 to 90

Why Does This Exist?

Sensors count photons, and dim scenes deliver few. Shot noise and read noise then rival the signal itself: night phone photos carry grain that no exposure slider removes, because the slider amplifies noise together with signal. Classical smoothing such as a Gaussian filter or a median filter blurs away the grain along with the detail it was meant to save.

Learned denoisers separate the two statistically. Noise is unpredictable from neighbouring pixels, while signal is. Networks trained on noisy-clean pairs learn that split at every brightness and texture, recovering edges that classical filters melt. Denoising also pays rent downstream: image enhancement stretches whatever range it is given, so cleaning first keeps the stretch from baking grain in permanently.

Think of It Like This

A sound engineer removing tape hiss

An old concert tape hisses under the music. The engineer cannot simply mute the high frequencies, because the cymbals live there too. Instead they sample a silent passage to learn the hiss fingerprint, then subtract that fingerprint everywhere, music preserved.

The noise model is the hiss fingerprint and residual learning is the subtraction. It fails on sounds the fingerprint misdescribes: a cymbal crash with hiss-like statistics gets subtracted as noise, which is the texture-loss limit below.

How It Actually Works

Model the capture as $y = x + n$ , where $y$ is the noisy capture, $x$ is the clean scene, and $n$ is Gaussian noise with standard deviation $\sigma$ . Residual networks such as DnCNN predict the noise $\hat{n} = f(y)$ and return $\hat{x} = y - \hat{n}$ , an easier target than the clean image because noise has simpler statistics than scenes. Blind denoisers handle a range of $\sigma$ values, while non-blind ones take $\sigma$ as input and specialise per level:

\hat{x} = y - f(y; \sigma), \qquad y = x + n

Burst methods average aligned frames for effectively noiseless ground truth, and self-supervised variants train on noisy pairs alone using the Noise2Noise insight that the mean of noisy targets is clean.

Worked example

Pixel scale $0$ to $255$ , noisy MSE $625$ ( $\sigma \approx 25$ ). PSNR is $10 \log_{10}(65{,}025/625) = 10 \log_{10}(104.04) = 20.17$ dB. After denoising, MSE $90$ : PSNR is $10 \log_{10}(65{,}025/90) = 10 \log_{10}(722.5) = 28.59$ dB. Error falls sevenfold and grain visibly dissolves while edges hold, the signature win of learned filtering over classical smoothing.

Code

Residual denoising arithmetic on one noisy pixel.

1def denoise_pixel(noisy: float, predicted_noise: float) -> float:2    """Subtract the predicted noise layer from the capture."""3    return noisy - predicted_noise4
5
6print(denoise_pixel(148.0, 23.0))  # -> 125.0 near the true value

Watch Out For

Over-smoothing fine texture into plastic

Aggressive denoisers erase skin pores, fabric weave, and foliage along with grain. Faces turn waxy at exactly the noise levels where users most want help.

Validate on textured crops at full zoom, not just smooth walls. Prefer denoisers with texture-preserving losses for portraits and nature.

Training noise that mismatches the sensor

Gaussian $\sigma = 25$ training meets Poissonian low-light sensor noise in deployment, and performance crumbles. Noise family matters more than network depth past a point.

Calibrate to the target sensor: measure its noise profile or train blind across families. Synthetic Gaussian benchmarks flatter mismatched models.

The Quick Version

Denoising models a capture as clean signal plus noise and predicts the noise layer.
Residual learning subtracts $\hat{n}$ from $y$ , an easier target than $x$ itself.
Our example climbs PSNR $20.17$ to $28.59$ dB as MSE falls $625$ to $90$ .
Burst averaging and Noise2Noise supply ground truth without clean captures.
Waxy texture loss and sensor-mismatched training noise are the standing traps.