Computer vision: an introduction for beginners

Your phone unlocks when it sees your face. A camera in a factory spots a scratch on a part in a fraction of a second, one a person would miss. A car recognizes a pedestrian at a crossing. All of this is computer vision: the field that teaches machines to understand images. Let’s look at how a computer “sees,” what tasks it can solve and where it is used.

What computer vision is

Computer vision is the set of methods for extracting useful information from photos or video. Not just storing a picture, but answering questions about it: what is in it? where exactly? how many? is there a defect?

For people, seeing is the easiest thing in the world. We recognize a friend in a crowd without thinking. For computers it was long one of the hardest problems. The reason becomes clear once you look at what a computer actually gets from a camera.

What a computer “sees”

An image is made of pixels, small squares. In a black-and-white image each pixel is one number from 0 (black) to 255 (white). A color image has three such tables: red, green and blue (RGB). An ordinary 12-megapixel camera produces 36 million numbers per frame.

So a computer doesn’t see “a washer.” It sees a table of numbers. The job of computer vision is to find patterns in those numbers that correspond to a washer, a scratch or a face, and to do it so the result doesn’t change when the lighting, angle or background shifts a little.

Three basic tasks

• Classification answers “what is in the image?” A good part or a defective one. A cat or a dog.
• Detection finds objects and draws boxes around them: “scratch here, 94% confidence.” This is how people are counted in a frame or boxes found on a pallet.
• Segmentation determines exactly which pixels belong to an object. You need it when the precise shape matters: the area of a stain, the outline of a part, the borders of a tumor on a scan.

Beyond these there are reading text and codes (OCR), measuring dimensions, tracking objects across frames and estimating 3D position.

Two approaches: classic and neural networks

Classic computer vision is a toolbox of proven algorithms: filters, edge and contour detection, template matching. The engineer decides what to look for: “if there is a dark spot larger than 2 mm on a light background, it’s a defect.” These algorithms are fast, predictable and work well in controlled conditions. The best-known library for them is OpenCV.

Deep learning changed the field from around 2012. Instead of describing features by hand, you show a neural network thousands of examples, and it learns by itself what separates a scratch from a reflection. Popular models such as the YOLO family run real-time detection even on small devices. For more on how neural networks learn, see the introduction to artificial intelligence.

In practice the two are often combined: classic methods prepare the image and take precise measurements, while a neural network handles the hard decisions.

How it works on a production line

1. Light and camera. Lighting is chosen to make the defect as visible as possible: for example, light at a low angle turns scratches into bright lines.
2. Capture. A sensor on the conveyor signals that the part is in frame, and the camera takes a picture.
3. Analysis. An algorithm or neural network processes the image.
4. Decision. Good or reject, what the dimensions are, where the part lies.
5. Action. The result goes to the PLC, which rejects the part or guides a robot.

The main lesson everyone who builds these systems learns quickly: good lighting matters more than a clever algorithm. If a defect is barely visible in the image, no neural network will find it reliably.

Where it is used

• Quality control: scratches, chips, wrong assembly, liquid level in a bottle.
• Reading codes and markings: barcodes, QR codes, dates on packaging.
• Robots that see: find a part lying any which way in a bin and pick it up.
• Safety: notice a person in a danger zone or a missing hard hat.
• Outside the factory: medical imaging, agriculture (crop health from drones), cars, checkout-free stores.

Where the data comes from

A neural network has to be trained, and that takes labeled data: thousands of images where a person has marked what is a defect and what isn’t. Factories have a paradox here: there are millions of good parts but only a handful of defective ones. So people use augmentation (artificially rotating, relighting and adding noise to existing images), synthetic data (generating realistic defect images) and models that learn to spot “something unusual” from examples of normal parts only.

Where the field is heading

• Vision foundation models. Large models trained on billions of images can segment almost any object without extra training. This sharply cuts the time it takes to start a project.
• Vision-language models. You can show them a photo and ask in text: “What’s wrong with this board?” For quick experiments this already works; for critical decisions on the line, caution is still needed.
• Edge AI. Neural networks run right inside smart cameras, without sending video to a server.
• 3D vision. Stereo cameras and structured light give not just a picture but depth.

How to get started

1. Install Python and the OpenCV library. Load a photo, convert it to grayscale, find the edges. That’s ten lines of code.
2. Take a ready-made pretrained detection model (for example from the YOLO family) and run it on your own photos or webcam video.
3. Try a small real problem: count the coins on a table, or tell a broken cookie from a whole one.

Computer vision is one of the most visible ways to see artificial intelligence at work in real production. In the next articles we will cover cameras, lighting and the first hands-on projects.