Computer vision is software that turns a grid of pixels into a label, a box, a mask, or a decision. A camera (or a scanner) captures the grid. The model learns which patterns in that grid mean “cat,” “car,” or “stop sign.” It is not a camera. It is not image processing, which outputs a cleaner image. It is the step after that — understanding.
Teach a machine what a cat is and you are not showing it one photo. You are showing it enough views — fluffy, leaping, cartoon — that the next unseen cat still matches. That is the job. Healthcare is one place this job is used. It is not the definition. The healthcare applications page is computer vision in healthcare.
What it means to see like a machine
To the computer, an image is a number grid. Each cell is brightness (and, in color, three channels). Nothing in that grid is already a cat. The model has to learn which combinations of numbers keep showing up when the label is “cat.”
Three skills fall out of that:
- Object identification. This patch is a car, not a lamppost.
- Scene understanding. This frame is a beach, not an intersection.
- Activity recognition. This sequence is a person walking, not standing still.
How a computer actually learns to see
Four steps. Same loop whether the sensor is a phone camera or a CT gantry.
- Acquisition. A sensor writes the grid. Garbage in, garbage out. Focus, exposure, protocol, and compression are the first decisions.
- Preprocessing. Denoise, fix contrast, resize, normalize so the model is not thrown by lighting or scale. That job has its own page: what is image preprocessing.
- Feature extraction. Edges, textures, shapes — or, in a modern net, features the net learned itself. You are compressing the grid into the clues that matter.
- Training, then inference. During training the model sees labeled examples and adjusts until the clues predict the label. At inference it sees a new image and returns a class, a box, or a mask. How you train the model is a different article: how to train a machine learning model.
The three techniques that actually ship
| Technique | Analogy | What it returns |
|---|---|---|
| Image classification | Sorting photos into albums — “beach” or “city” | One label for the whole image |
| Object detection | Drawing boxes around every car in a street photo | A class plus a location (the box) for each object |
| Image segmentation | Coloring a person in, pixel by pixel | A mask — the exact shape, not a rectangle |
Classification answers “what is this image?” It does not say where. Sorting factory parts, tagging an X-ray as “chest” vs “hand,” or “day” vs “night” are this job.
Detection answers “what is in this image, and where?” Bounding boxes on cars, people, stop signs. That is what a driving stack uses to keep the car in the lane and off the pedestrian.
Segmentation traces the outline. Semantic segmentation paints every pixel with a class (all “road” pixels, all “sky” pixels). Instance segmentation also splits two people into two masks. In medical imaging this is the organ or lesion outline. The healthcare page covers that use; the technique itself is not a clinic-only idea.
Healthcare is one job, not the whole field
The same three techniques read a scan. Classification grades a fundus photo. Detection flags a polyp or a nodule. Segmentation outlines a tumor so someone can measure volume. That is useful, and it is still assistive: the clinician signs. The applications, the “will it replace clinicians” question, and how a clinic actually starts live on computer vision in healthcare — we already pointed there, and that is the canonical page for the clinic job.
Applications beyond the clinic
- Driving. Cameras plus detection and segmentation: other cars, people, signs, lane paint. Real time, or the car is a paperweight.
- Retail. Overhead cameras add items to a virtual cart. The same stack watches shelves for empty facings.
- Manufacturing. A camera over a line flags a crack, a misaligned label, a color miss at a rate a person cannot hold for a shift.
- Agriculture. A drone pass plus a multispectral camera marks stressed canopy so water or spray goes only where the plants need it.
Questions that keep coming up
Computer vision vs image processing
Image processing outputs another image — sharper, less noisy, higher contrast. Computer vision outputs a decision: a label, coordinates, a mask. Processing is often the cleanup step before vision. It is not the same job.
Is computer vision the same as AI?
No. Artificial intelligence is the umbrella (learn, decide, act). Computer vision is the sight piece under that umbrella. A model that reads a table of labs is AI without vision. A model that boxes a car is vision.
Does it always need millions of labeled images?
Not anymore. Transfer learning starts from weights already trained on a large general set (ImageNet is the usual example) and fine-tunes on your smaller set. Augmentation (flip, rotate, re-light) stretches a small set. Synthetic images fill gaps when real ones are rare or too sensitive. You still need honest labels for the job you want. You do not always need a million of them.
How do you start?
Python. Then OpenCV for loading and simple ops. Then PyTorch or TensorFlow for a first net. Pick a small, boring project (flowers, a face, a stop sign) and finish it. Tutorials are cheap. A finished loop is the lesson.
PYCAD uses this stack on medical images — custom pipelines and web DICOM viewers when the study has to live in a clinic app. The healthcare write-up is here. Case studies.