Deblurring

A photo becomes blurry when the camera shakes during the shot or the subject moves. Each pixel is then smeared across several neighboring pixels. Deblurring is the attempt to reverse this smearing and recover the sharp image. This is harder than it sounds: the blurred image alone does not unambiguously reveal what the original looked like. A program must therefore guess which sharp version is most plausible. It is precisely this guessing that is now handled by programs that have learned from a great many example images what real images normally look like.

Why blurry images are more than just an annoyance

For personal photos, blur is merely unattractive. In other areas it renders images unusable. A surveillance camera that captures a license plate only as a smear delivers no usable information. A medical microscope image in which cell boundaries blur cannot be reliably evaluated.

Deblurring is also a substitute for expensive hardware. Sharp images in low light would normally require large lenses and highly light-sensitive sensors. A phone has no room for that. Instead, it captures dark images with a long exposure time and corrects the blur afterward via software. This principle is called computational photography: computing power replaces optics.

An important distinction is needed here. Deblurring removes blur, super-resolution increases resolution, and denoising removes image noise, i.e. the grainy color speckles that appear in darkness. The three tasks are related and are often solved jointly by a single model. But each refers to something different.

From the blur formula to the learning network

Classically, blur is described mathematically using a so-called kernel. This is a small table of numbers indicating how strongly a pixel is smeared onto its neighbors. If this kernel is known, the calculation can in principle be reversed. With motion blur, however, the kernel is unknown, because no one recorded the exact hand movement. The method must then estimate the kernel and the sharp image simultaneously.

Modern systems bypass this problem. They train a neural network, that is, a program that derives its computational rules itself from examples. To do this, one takes millions of sharp photos and artificially blurs them. Each pair of blurred and sharp image is a training example. The network adjusts its internal numerical values until its output comes as close as possible to the sharp original.

A common misconception is that lost information is thereby recovered. That does not happen. The network invents details that are statistically consistent with what it has learned. Often it guesses well, but sometimes it generates letters or facial features that were never present in the original. In court, images reconstructed in this way are therefore not reliable evidence.

Deblurring in phone cameras and image archives

Deblurring is most often encountered without being noticed: in the smartphone’s camera app. Features such as night mode or face sharpening combine several shots and compute away the motion blur. Image editing programs also offer tools with names like “sharpen” or “unsharp mask.” Some merely enhance edges, others use genuine deblurring models.

In the news, the topic comes up in camera launches and in disputes over photographic evidence. When manufacturers advertise “AI photography,” deblurring is usually part of it. At the same time, experts debate at what point a photo edited in this way ceases to be a document of reality.

Deblurring is also technically relevant outside of photography. Astronomers compute away the distortion caused by atmospheric layers from telescope images. Archives sharpen old film footage for restoration purposes. And self-driving cars need sharp images at high speed, since otherwise their cameras only see streaks.