Applied AI · A visual guide

What is computer vision?

It is an area of artificial intelligence that analyses images or video to find useful visual signals for a specific task.

Explore the guide
OBJECT · 0.92 OUTPUT HUMAN REVIEW
Conceptual examples with illustrative images; they do not represent T2G results or specific projects.

In simple terms

A computer can look for visual patterns in an image.

A system receives images, extracts features and compares them with patterns learned or rules defined for a task. It can then classify an image, locate an object or identify a region. The output is an estimate that must be assessed in its actual use context.

At T2G: we start with the decision to support and the conditions of the available images. We then assess which vision task makes sense, how to measure it and where human review is needed.

How a vision solution is built

From an image to a signal you can evaluate.

The process depends on the task. These steps help define what a system should recognise and check how it responds.

  1. 01

    Define the task

    Specify which object, condition or region matters and what a person would do with the result.

    THE VISUAL QUESTION INPUT IMAGE TARGET OBJECT
    Conceptual schematic01 / 04
  2. 02

    Prepare examples

    Gather relevant images, check their quality and label the examples needed to teach or assess the task.

    EXAMPLES IN CONTEXT DIFFERENT BACKGROUNDS · ANGLES · CONDITIONS
    Conceptual schematic02 / 04
  3. 03

    Evaluate in context

    Measure correct and incorrect results on separate images, including operating variations in light, angle, background or camera.

    REAL CONDITIONS LIGHT ANGLE BACKGROUND THE SAME TASK CAN PRODUCE DIFFERENT ERRORS
    Conceptual schematic03 / 04
  4. 04

    Integrate carefully

    Decide how to present the output, which cases need review and how to detect image changes over time.

    FROM MODEL TO DECISION MODEL ? OUTPUT REVIEW UNCERTAIN CASES FOLLOW A REVIEW PATH
    Conceptual schematic04 / 04

Three common tasks

The output depends on what you need to know.

CLASSIFICATION INPUT IMAGE CATEGORY
01

Classification

Assigns a category to a whole image, such as whether it meets a visual criterion.

DETECTION INPUT IMAGE LOCATION
02

Detection

Locates objects and marks where they appear with boxes, supporting counting or inspection.

SEGMENTATION INPUT IMAGE REGION
03

Segmentation

Marks which pixels belong to a region when shapes or surfaces need to be outlined.

Judgement before automation

The environment matters as much as the model.

  • 01

    Images unlike those used for training can increase errors.

  • 02

    A detection is an estimate; uncertain cases may need expert review.

  • 03

    Before automating, agree on which errors are acceptable for the task.

Next step

Explore the service and keep learning.

Talk about your challenge

Could images help solve a task in your operation?

We can help define the question, data and criteria before choosing a computer vision approach.

Book a diagnostic session directly with the partners