Quick answer: computer vision is the branch of artificial intelligence that lets software make sense of images and video, then act on what it finds. A trained model looks at pixels and returns a label, a location, or a decision, without a person marking up every frame by hand. You meet it whenever a phone unlocks by recognizing your face, a store camera tracks a shelf, or a car reads a stop sign.
This entry covers what computer vision actually does, how the pipeline from camera to decision works, where you already meet it, how it differs from plain image processing, and where the technology still breaks.
What Is Computer Vision?
Computer vision is the artificial intelligence discipline that gives machines the ability to extract meaning from visual input: photos, video frames, or a live camera feed. The goal is not to reproduce human eyesight but to turn pixels into something a program can act on, such as “this is a pedestrian” or “this shelf slot is empty.”
That distinction matters. A camera captures light. Computer vision decides what the light means.
The field is decades old, but it stayed largely theoretical until deep learning caught up with it. In 2012, a neural network called AlexNet beat every hand-engineered approach at the ImageNet image recognition competition, and the field has been dominated by trained models ever since.
Most computer vision work now falls into a handful of tasks: classification, which asks what is in an image; detection, which asks where it is and draws a box around it; and segmentation, which marks the exact pixels that belong to it. A single product often chains two or three of these together, running detection first to find a face in a photo, then a separate model to identify whose face it is.
Video adds a further wrinkle: tracking, which follows the same object across many frames instead of scoring each frame on its own. That is what lets a security camera flag a person walking through a restricted area rather than just noting that a person exists in one still image.
How Does Computer Vision Work?
The pipeline starts with a sensor. A camera captures light and converts it into a grid of pixel values, the same raw material a photo editor works with.
That raw image usually gets cleaned up first: resized, adjusted for brightness, stripped of noise. Then it passes into a neural network, most often a convolutional neural network, which slides small filters across the image and learns to notice edges, then textures, then shapes, then whole objects, one layer at a time.
Newer systems increasingly use a different architecture called a vision transformer, borrowed from the large language model world, which treats an image as a grid of patches rather than a single sliding window. Both approaches train the same basic way: shown millions of labeled examples, scored on how wrong each guess is, and nudged toward fewer mistakes.
That last part is the catch. A model is only as good as the labeled images it trained on, and getting people to draw boxes around millions of photos is slow and expensive. That is why teams increasingly fill gaps in real footage with synthetic data: simulated scenes built to cover situations a camera rarely captures on its own, like a child running into the road at night.
Where Is Computer Vision Used?
Self-driving systems lean on computer vision to find lane lines, read traffic signs, and spot pedestrians in real time. Waymo’s robotaxis, which reported carrying around 500,000 paid rides a week in early 2026, run on exactly this kind of perception stack alongside radar and lidar.
Medical imaging is where the technology has moved fastest into daily use. Radiologists now work alongside AI tools that flag suspicious regions on X-rays, CT scans, and mammograms; by 2025 the FDA had cleared more than 1,400 AI-enabled medical devices, most of them for reading radiology images.
Retail uses it for cashierless checkout and shelf monitoring, spotting what is out of stock without a person walking the aisle. Farms use it too: cameras mounted on tractors identify individual weeds so a sprayer can target only those plants instead of an entire field.
Your own phone runs a version of it every time face unlock checks who is holding it, and every camera app that blurs a background or merges a burst of shots into one clean photo is doing the same kind of work.
Computer Vision vs Image Processing: What’s the Difference?
The two get used interchangeably, and they are not the same step.
Image processing changes an image without understanding what is in it: sharpening, adjusting contrast, removing noise. A photo filter is image processing.
Computer vision interprets an image and produces a decision or a label. It often uses image processing first, as cleanup, before a model tries to understand the result. A closely related term, machine vision, refers to computer vision purpose-built for a factory line: fixed cameras checking parts for defects at high speed, a narrower and more predictable job than general-purpose vision.
What Are the Limits of Computer Vision?
Models trained on one set of conditions struggle when the world drifts from that training data. US safety regulators opened an investigation into Tesla’s Full Self-Driving in October 2024 after crashes in reduced visibility, including sun glare, fog, and airborne dust, and one of those crashes killed a pedestrian.
Glare and fog are exactly the conditions a camera-only system has least margin for, and exactly the conditions that show up least often in training footage.
Small objects are a known weak point too: a convolutional network works from limited pixel detail, so a distant pedestrian or a small defect on a production line is easy to miss.
Adversarial examples make the problem worse on purpose. Researchers have shown that tiny, human-invisible changes to an image can flip a model’s answer entirely, evidence that these systems recognize statistical patterns rather than objects the way a person does.
Bias is the quieter risk. A face recognition system trained mostly on one demographic performs worse on the ones it saw least, which is why several US cities and companies have restricted or paused police use of the technology.
None of that makes computer vision unreliable for its actual job. It makes it a tool that needs a known operating range, careful testing at the edges, and a person in the loop wherever a wrong answer costs something.
Where Is Computer Vision Headed?
The clearest trend is convergence with language. Vision-language models now answer questions about an image in plain sentences instead of returning a fixed label, which is why a growing share of image tools double as chatbots that can also see.
The other trend is smaller and cheaper models running directly on a phone or a camera rather than a server, since sending every frame to the cloud is slow and raises its own privacy questions. For a reader picking tools in 2026, that shift is worth watching: on-device vision is where the practical gains are landing first.