Computer Vision

To a computer, a photo initially consists of nothing but millions of numbers. Each pixel stores how much red, green, and blue it contains. A human looking at the same image immediately sees a dog on a lawn. Computer vision is the field that aims to close this gap. It develops programs that turn those numbers into a statement about the image’s content. Such statements can include: What objects are visible? Where exactly are they located? And what is happening over the course of a video?

Why machines should learn to see

A great many tasks in business consist of looking at images and making judgments about them. A worker inspects parts on an assembly line for scratches. A doctor examines an X-ray for suspicious spots. A security service monitors cameras. This work is monotonous, and attention wanes after hours on the job. A machine does not get tired and inspects every part by the same standard.

Then there is the sheer volume of visual material. Billions of photos are taken worldwide every day, and countless cameras run around the clock. No human could even begin to review this data. Only automatic analysis makes it usable at all. That is precisely why computer vision is one of the most economically important areas of artificial intelligence.

The field also has a critical side, however. The same technology that finds lost pets in photos can identify people in public spaces. That is why the EU’s AI Act strictly regulates the use of facial recognition. Anyone talking about computer vision is almost always talking about data protection as well.

From pixels to meaning

Early systems worked with hand-programmed rules. A developer would describe, for instance, what an edge looks like: the place where light and dark pixels meet. Edges were meant to form shapes, and shapes were meant to form objects. This worked in laboratory settings but failed in the real world. A chair in backlight looks completely different from the same chair in daylight.

Today, systems learn from examples rather than rules. A so-called neural network — a computational model capable of learning — is shown hundreds of thousands of labeled images. For each image, it compares its guess with the correct answer and adjusts its internal numerical values by a tiny amount. After millions of such passes, the guess is usually correct. At no point was the network ever told what a dog looks like.

Inside, the network operates in layers. The first layers respond to very simple patterns such as edges and color gradients. Later layers combine these into eyes, wheels, or letters. The final layer delivers the verdict. Think of it like an administrative office: details are recorded at the bottom, and decisions are made at the top.

Cameras that already think along with you

The next example is probably sitting in your pocket right now. The photo gallery on your smartphone sorts pictures by people and places. The camera app recognizes faces and focuses on them. When paying, the device unlocks itself via facial scan. All of this is computer vision, even though the term never appears anywhere.

In industry, cameras inspect welds and packaging several times per second. Cars recognize traffic signs, lane markings, and pedestrians. Agricultural machines distinguish crops from weeds and spray herbicide only in targeted spots. In hospitals, software flags suspicious areas on scans before a doctor evaluates them.

In business news, the term usually comes up in connection with chip manufacturers, automakers, or factory-automation start-ups. One distinction is important here: image generators like Midjourney create images, while computer vision understands them. Both work with similar technology but pursue opposite directions. Newer models are now capable of both and are therefore called multimodal.