Learn AI /Computer Vision and Language Technology /Computer vision

Computer visionCV

Also called: machine vision.

Detail

In five seconds

What it is

Computer vision enables software to analyse images or video and produce useful outputs such as labels, locations, measurements or text.

In plain words

Computer vision helps software pull useful information from pictures or video, such as recognising an object, finding where it is, or reading text in an image.

An everyday example

At a busy train station, staff receive phone photos of items that travellers have lost: a blue suitcase, a red scarf, a passport cover and a pair of glasses. One worker looks at each photo and writes useful notes: what the item seems to be, where it appears in the picture, and any visible text or distinguishing marks. Another worker uses those notes to place the item in the right recovery queue. The staff are not reading minds or knowing the full story of the object; they are extracting practical clues from what is visible.

Where the story stops working

A human worker can use broad life experience and ask follow-up questions. A computer vision system only uses patterns it was built or trained to use, and its output quality depends heavily on the data, camera conditions and task definition. Also, many systems do not first make human-like notes and then act; they may produce task outputs directly through mathematical processing.

Remember thisComputer vision is not one tool but a family of methods that turn images or video into task-specific outputs.

Understand it

Why it matters

Non-specialists often encounter claims that a system can 'see'. That phrase can hide important differences. A tool that reads printed text from scans is not the same as one that identifies faces, counts vehicles or maps flooded areas. Knowing the term helps you ask better questions: What exactly is the task? What images were used to train it? How accurate is it under poor lighting, unusual angles or different locations? What are the privacy, bias and safety risks? Those questions matter when buying tools, setting rules, checking evidence and deciding how much human oversight is needed.

How computer vision turns an image into a task output

Show that computer vision takes visual input, analyses pixel patterns for a defined task, and returns different kinds of outputs depending on the goal.

01
1. Visual input
The process starts with a photo, scanned document, medical image, satellite image or video frame.
Draw a box with small icons for a camera image, a scanned page and a video frame. Label the box 'image or video input'.
02
2. Pixels as numbers
The image is represented as pixels, which are tiny picture units stored as numbers for colour or brightness.
Draw a zoomed-in grid beside the first box, with a few cells shaded differently and small numbers shown in some cells.
03
3. Pattern analysis
A model analyses combinations of pixels to find useful visual patterns such as edges, shapes, textures or character strokes.
Draw a box with layered filters or simple shapes highlighting an edge, a rectangle and letter-like marks. Use arrows from the pixel grid into this box.
04
4. Task choice
The same image can be analysed for different tasks: classify the whole image, detect objects, segment regions, read text or estimate pose.
Draw a branching box with five labelled mini-panels: 'classify', 'detect', 'segment', 'read text', 'estimate pose'.
05
5. Output for use
The system returns an output that another process or a human can use, such as a label, count, highlighted region or extracted text.
Draw a final box showing example outputs: 'damaged road', '3 vehicles', a highlighted area and a short line of extracted text.

A left-to-right flow with five boxes connected by arrows. Read from the incoming image or video, through pixel data and pattern analysis, to task-specific outputs such as labels, object locations, extracted text or measurements. The final note reminds the reader that the task shapes the output.

Hovering or tapping each step highlights the matching visual element and shows one sentence of explanation; arrows animate only if motion is allowed.

Describe this without the diagram

This diagram explains computer vision as a five-step flow. First, the input is an image or video, such as a photo, scanned page or live frame. Second, the image is stored as pixels, meaning many small picture units represented by numbers. Third, a model analyses patterns across those pixels, such as edges, shapes, textures or letter strokes. Fourth, the task determines what kind of analysis matters: whole-image classification, object detection, image segmentation, optical character recognition or pose estimation. Fifth, the system outputs something usable, such as a label, a count, a marked location, a highlighted region or extracted text. The main idea is that the same visual input can lead to different outputs depending on the task.

With reduced motion: A static left-to-right diagram with the same five boxes and arrows, with no animation. Each box remains visible at once, and the task-choice box shows all example tasks side by side.

Computer vision is a broad area of AI that works with images and video. The input might be a photo, a scanned page, a camera feed, or a satellite image. The output depends on the task: a label such as “road damaged”, a marked location around an object, extracted text, or a count.

A system usually starts by turning the picture into numbers. The smallest picture unit is a pixel. Patterns across many pixels can suggest edges, shapes, colours or textures. For many tasks, a machine learning model is trained on examples so it can link visual patterns to a task, such as spotting vehicles, reading printed words, or separating land from water.

What happens next depends on the goal. Some systems classify the whole image into one category. Others detect several objects and show where they are. Others divide an image into regions, estimate body pose, or describe an image in words. So computer vision is the wider field, not one single method.

People often confuse computer vision with facial recognition or image generation. Facial recognition is one narrower use that tries to match or verify faces, and it raises distinct privacy and rights concerns. Image generation creates new images. Computer vision usually analyses existing visual input to extract information, though some modern systems combine both kinds of capability.

How “A lost-and-found desk sorting photos” maps onto the real thing

In the story	In the AI system
staff receiving phone photos	a system taking image or video input
worker writing notes about item type, location and visible text	a model producing outputs such as labels, bounding boxes and extracted text
placing the item in the right recovery queue	using visual analysis for a downstream task such as routing, counting or search
distinguishing marks in the photo	visual features in pixels that help the model separate one pattern from another

A worked example

A conference services assistant at Headquarters receives a paper statement after a Main Committee meeting and must prepare it for upload to the portal before close of business.

Starting from: A scanned page containing a heading, several paragraphs and a footer.

The scanner represents the page as an array of pixel values.
A layout model locates the heading, body text and footer and determines their reading order.
Optical character recognition maps the shapes in each text region to characters and words, assigning confidence scores to its predictions.
Words with low confidence are highlighted for the assistant to compare with the scanned page and correct.

Result: The assistant obtains searchable text alongside the original scan, corrects the flagged words and uploads the checked version to the portal.

In a United Nations workplace

In a Resident Coordinator's Office, a budget officer has until close of business to clear a response to an audit observation on the asset register. The file includes photographs of equipment labels submitted by premises focal points, which must be checked against the register before the response goes to the front office.

An approved computer vision system could locate asset labels in the photographs, extract the printed tag numbers and identify broad equipment types. It would produce structured fields for reconciliation, leaving the budget officer to confirm each match against the original image and the register.

Before you try thisPhotographs may reveal staff, access badges, screen contents, room layouts or location metadata. They must not be submitted to a tool that is not approved for such material. Crop working copies to the label where possible, remove unnecessary metadata, restrict access and verify every extracted tag before using it in the audit response.

Where you may hear it

A vendor demo to a procurement unit assessing document-scanning software
An information management meeting on extracting text from scanned situation reports
An ICT governance board reviewing image-analysis tools and data safeguards
An internal audit discussion on verifying asset records from photographs

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Computer vision means a system sees images the way people do.
Correction: The name is convenient, but the mechanism is different. The system analyses numerical patterns in pixels for a defined task. It may perform well on one narrow task and still fail on scenes that are easy for a person, especially when lighting, angle or context changes.
Common misconception: It is basically the same as facial recognition.
Correction: Facial recognition is only one subfield within computer vision, and it is a particularly sensitive one. Many computer vision systems never analyse faces at all; they may read documents, detect crops in satellite images or find defects in manufactured parts.
Common misconception: If the output looks precise, it must be correct.
Correction: Visual outputs such as boxes, labels and confidence scores can look authoritative even when the underlying result is wrong. Errors can come from biased training data, blurry images, unusual environments or a mismatch between the training task and the real task.
Common misconception: Computer vision always needs live cameras watching people.
Correction: Many uses involve still images or scans rather than surveillance. Examples include reading printed text on archived documents, detecting cracks in bridges, or mapping storm damage from satellite images. The privacy risk depends on the data and purpose, not only on the term itself.

Quick check

Which example best matches computer vision rather than a different AI area?

Go deeper

The simple version

Computer vision helps software pull useful information from pictures or video, such as recognising an object, finding where it is, or reading text in an image.

The technical version

Computer vision is the field concerned with computational methods that map visual input to structured outputs. The input may be a single image, a video sequence or another image-like signal. The system represents the input as arrays of pixel values and uses algorithms or learned models to extract patterns relevant to a task. Those tasks include image classification, object detection, image segmentation, optical character recognition, pose estimation and image captioning. In many current systems, the core model is trained on labelled examples so that parameter values are adjusted to reduce error on the target task. Performance depends on data quality, coverage, annotation choices, camera conditions and evaluation methods. The field overlaps with image processing, but image processing often focuses on transforming images, while computer vision focuses on inferring information from them.

Reviewed external resources

EncyclopaediaComputer vision (opens in a new tab)Wikipedia
This link gives a broad overview of the field and suits readers who want a general reference with many related topics.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This glossary helps readers check short definitions of technical terms they may meet while learning about computer vision and suits quick look-ups.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course introduces machine learning concepts that support many computer vision systems and suits beginners who want background before going deeper.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Computer vision (opens in a new tab)YouTube search
This is a YouTube search results page, so it helps readers find video explanations in different styles and suits people who learn best by watching.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Computer vision (opens in a new tab)Reddit search
This is a Reddit search results page, so it helps readers see community questions and discussions and suits readers who want informal explanations and common confusions.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

Computer vision is not one tool but a family of methods that turn images or video into task-specific outputs.