Computer vision in practice

What it isCopy Anchor Link

The main jobs are classification (what is in this image), detection (where are the objects, drawn as boxes), segmentation (which pixels belong to which object) and reading text (OCR). Modern systems start from a model pretrained on millions of images and fine-tune it on a few hundred of your own, which makes small projects possible. Vision-language models can also answer questions about an image in plain words, which is often enough for simple cases.

Where it is usedCopy Anchor Link

Factories check parts for defects on a conveyor belt. Farms spot diseased plants from drone photos. Hospitals use models that flag suspicious areas on scans for a doctor to review. Banks read cheques and invoices. Phones unlock with a face. In each case a human usually reviews the uncertain cases.

A worked example: a screening testCopy Anchor Link

A hospital screens 1,000 scans. Ten show disease. The model finds 9 of the 10 (recall 90 percent) and raises 50 false alarms. It flagged 59 scans, of which 9 are real, so precision is 9 / 59, about 15 percent. That can still be a good screening tool, because a doctor reviews 59 scans instead of 1,000 and only 1 case was missed. The numbers decide whether it helps, not the headline accuracy.

What goes wrongCopy Anchor Link

Models learn shortcuts. A classifier trained on photos from one hospital may be reading the scanner's label, not the disease. A model trained in good light fails at night. Always test on images from a different place and time than the training set, and keep checking after launch because cameras and conditions change.

How a detection gets scoredCopy Anchor Link

A detector draws a box. To grade it, compare with the true box using intersection over union, or IoU: the overlap area divided by the combined area. Say the true box covers x from 0 to 10 and y from 0 to 10, and the prediction covers x from 5 to 15 and y from 0 to 10. The overlap is 5 × 10 = 50. The union is 100 + 100 - 50 = 150. IoU is 50 / 150 = 0.33. Teams choose a cutoff, often 0.5, and count a box as correct only above it. This is why a box that "looks about right" can still count as a miss.

Routing by confidenceCopy Anchor Link

A model gives each image a score between 0 and 1. Invented example with 1,000 images: 800 score above 0.9 and pass automatically, of which 4 are wrong (0.5 percent). The other 200 go to a person, and 30 of them are wrong, which the person fixes. People handle 20 percent of the load, and, if the reviewer catches every error, the final error rate is 0.4 percent. The catch is calibration: whether a score of 0.9 really means 90 percent right. To check, group past predictions by score and compare each group's true accuracy with its score. Models are often too sure of themselves.

When a simple sensor winsCopy Anchor Link

A packing line needs to know whether a box has a bottle in it. A weight sensor or a light beam answers in a fraction of a millisecond, costs little and does not care about lighting. Use vision when the question depends on appearance, such as a scratch or a label. If a cheaper signal answers the question, use that.

Common mistakeCopy Anchor Link

Training and testing on images from the same camera session. The score looks great and falls apart in the real world.