What is computer vision?

Computer vision is the part of AI concerned with getting useful information out of images and video. A camera gives you a grid of coloured pixels. A computer vision model turns that grid into labels, boxes, outlines, counts or measurements that other software can act on.

Most business uses fall into a few task types:

• Classification: what is this whole image? (good part or defective part)
• Object detection: where are the objects, and what are they? (a box around each person, vehicle or carton)
• Segmentation: which exact pixels belong to each object? (the outline of a crack or a wound)
• Tracking: the same object followed across video frames, so you can count things crossing a line.
• Text reading: OCR is computer vision applied to printed or handwritten characters.

Modern computer vision uses neural networks trained on labelled images. During training the model sees thousands of examples ("this box is a helmet, this one is a bare head") and adjusts its weights until its predictions match the labels. Popular families include convolutional networks such as the YOLO detectors and vision transformers. Libraries such as OpenCV handle the plumbing: reading camera streams, resizing, cropping and drawing results.

At run time the model does inference on each frame. On a server that usually means a GPU; on a shop floor it often means a small edge AI device next to the camera, so video never has to leave the building.

A packaging unit wants to stop boxes leaving without a printed batch code. A camera above the line captures each box. A small detection model, trained on a few hundred photos of boxes with and without codes, runs on an edge device beside the conveyor. When a box has no code, the device sends a signal to a reject arm and logs a photo. A supervisor reviews the logged photos weekly, and false rejects become new training images.