What is computer vision?
Computer vision is the branch of artificial intelligence that teaches software to interpret images and video. A computer vision system takes pixels as input and returns something a business can act on: a label (“this is a damaged carton”), a location (“the forklift is in aisle 4”), a measurement (“the weld is 2 millimeters short”), or a line of text read from a shipping label.
Modern computer vision is built on deep learning. A convolutional neural network (CNN) or a vision transformer learns, from thousands of labeled example images, which patterns of pixels correspond to which outcomes. Older systems relied on hand-written rules about edges, colors, and shapes. Those rules still serve well in controlled settings such as a fixed camera over a conveyor, and many production systems combine both approaches.
How computer vision works
A production vision system has four stages. Cameras or scanners capture images under consistent lighting. Preprocessing crops, resizes, and normalizes the images. A trained model produces predictions with confidence scores. Business logic then decides what happens with each prediction, such as rejecting a part, opening a ticket, or routing a document to a queue.
The model is the smallest part of the project. Most of the effort goes into collecting representative images, labeling them accurately (see data annotation), and handling the cases the model gets wrong. A model trained on daytime photos of a loading dock will struggle at night unless night images were in the training data, and a model trained under one plant’s lighting may need new images at the next plant.
Common computer vision techniques
The main tasks differ in the question they answer and in the labeling effort they require. Labeling cost rises sharply from the top of this table to the bottom.
| Task | What the model returns | Typical business use |
|---|---|---|
| Image classification | One label for the whole image | Sorting incoming documents by type |
| Object detection | A box and a label for each object found | Counting pallets on a dock |
| Image segmentation | A class for every pixel | Measuring the exact area of a defect |
| Optical character recognition | The text that appears in the image | Reading invoice numbers and serial plates |
Object tracking, pose estimation, visual search, and facial recognition build on these foundations. Vision-language models now perform several of these tasks from a plain-language instruction with no task-specific training, at a much higher cost per image.
Computer vision in practice
A third-party logistics provider mounts cameras at its dock doors. An object detection model counts pallets as they are unloaded and reads the shipping labels, and the counts are matched against the advance shipping notice in the warehouse management system. Mismatches surface within minutes instead of at the end of the shift. The same footage feeds a safety model that flags people walking inside the forklift lane.
In a hospital, a segmentation model outlines the boundary of a wound in a photo taken on a tablet, and the measured area is written to the patient record so clinicians can track healing between visits. In a retail chain, a shelf camera compares the current shelf with the planogram and alerts staff to empty facings. In a county assessor’s office, a classifier sorts scanned permit applications by form type before a clerk sees them.
What to watch for
- Lighting, camera angle, and lens quality decide more of the outcome than the choice of model, so fix the capture setup before training anything.
- Ask how many labeled images a vendor needs and who will label them, because labeling is usually the largest cost in a vision project.
- Measure precision and recall separately on your own images, since a demonstration on the vendor’s images says little about your conditions.
- Decide where inference runs. Edge AI on a device next to the camera avoids streaming video to a data center and keeps latency low.
- Any camera that can see people raises privacy and consent obligations, even when the model never identifies anyone.
Computer vision and private AI
Video is heavy and often sensitive. Streaming every frame to a cloud service is expensive and may conflict with data residency rules or customer contracts. Many deployments therefore run inference on premises or at the edge and send only events and summary data upstream. Open-weight vision models make this practical for mid-sized companies. The same governance that applies to language models applies here: access control on the footage, audit logs of what the system flagged and who acted on it, and a human decision wherever a prediction changes a system of record.