AI · LESSON 14 / 25 · English
Computer Vision
An image becomes numbers before it becomes a prediction.
Content reviewed: 22 October 2026 · தமிழில் படிக்க ↗
Understand it in plain language
Computer vision works with images or video for tasks such as classification, detection, segmentation, and document reading. Pixels are only the starting representation. Models may use convolutional networks, Transformers, or other architectures. A system tested on neat images can fail under blur, lighting changes, unfamiliar cameras, or a different environment.
How the AI approach works
Computer vision maps visual inputs to task-specific outputs. Classification assigns labels to an image; detection identifies objects and their locations; segmentation assigns labels at pixel level. OCR extracts text, and visual tracking follows objects over time. Choose the output before collecting labels: image-level labels cannot directly supervise precise object boxes without additional methods or assumptions.
Images become tensors with spatial and channel dimensions. Preprocessing may resize and normalize them, but must preserve information important to the task. Convolutional networks learn shared local filters; vision Transformers process representations such as image patches with attention. A pretrained backbone can provide features, while a task-specific head produces the required output.
Visual shortcuts are a major risk. A model may learn a watermark, background, ruler, or camera artifact correlated with labels instead of the intended object. Split related captures together and test on different acquisition conditions. Augmentation should represent plausible variation without changing the label; horizontal flips are inappropriate when left–right orientation defines the answer.
Evaluate at the task level: classification metrics for image labels, overlap and detection criteria for boxes or masks, and text error measures for OCR. Include lighting, blur, occlusion, unfamiliar devices, and important subgroups. A confidence score cannot establish that a class absent from training has been recognized correctly.
Break the idea into steps
- Define whether you need a label, bounding boxes, pixel masks, or text.
- Collect varied images and keep related images together when splitting.
- Evaluate realistic errors and how uncertainty reaches the user.
Worked example: connect the method to the task
A shop wants to count damaged packages on a conveyor. Whole-image classification only says whether some damage is present; it does not count individual packages. Detection with one box and damage label per package better matches the requirement. If damage boundaries themselves matter, segmentation may be needed.
Collect examples from several shifts and cameras. Keep frames from the same video together when splitting. Review missed small defects and false alarms from shadows. The correct task formulation matters before architecture selection.
Apply the idea: check your reasoning
Think before revealing the answer
Why is average image brightness insufficient for recognizing a damaged package?
Reveal the answer and reasoning
It discards spatial structure and can change with lighting even when the package is unchanged. The task needs relevant visual features and labelled evaluation, not a brightness threshold presented as recognition.
A common mistake: Do not split near-identical frames of one video across training and test sets.
Before you move on
- Explain the main idea without repeating the technical label.
- Identify the approach, its assumptions, and the evidence needed to trust the result.
- Complete the challenge, then compare your reasoning with the answer.
Sources & a deeper next step
The refresh uses primary papers and official documentation. These optional links are in English; you can understand this lesson without reading them first. Product versions and availability can change.