Many business processes still depend on someone looking at something: inspecting parts on a production line, checking that shelves are stocked, reading meter dials, confirming a parcel was delivered undamaged. Computer vision is the field of AI that lets software interpret images and video. It has become far more capable and affordable in recent years, but it works best on narrow, repetitive visual tasks in controlled conditions. This article explains the main techniques in plain language, walks through realistic computer vision use cases by industry, and covers the practical requirements and limits.
What computer vision can actually do
Most business applications combine a handful of core tasks:
| Task | What it answers | Example |
|---|---|---|
| Image classification | What is in this picture? | Is this part acceptable or defective? |
| Object detection | What objects are where? | Draw boxes around each carton on a pallet |
| Segmentation | Exactly which pixels belong to what? | Outline the area of a scratch or stain |
| Optical character recognition (OCR) | What text is visible? | Read a serial number, label or number plate |
| Counting and tracking | How many, and how do they move? | Count items passing on a conveyor |
| Anomaly detection | Does this look different from normal? | Flag unusual surfaces without examples of every defect type |
Newer multimodal AI models can also describe images and answer questions about them in natural language. They are useful for flexible, low-volume tasks, but for high-volume, high-accuracy checks a specialised model trained on your images is usually faster, cheaper per image and more consistent.
Computer vision use cases by industry
Manufacturing
- Visual quality inspection: spotting cracks, scratches, missing components, misaligned labels or incorrect assembly. Cameras check every item rather than a sample, and do not tire on a night shift.
- Assembly verification: confirming the right part was fitted in the right place before the next step.
- Safety monitoring: detecting whether protective equipment is worn in designated zones, or whether someone has entered a restricted area near machinery. This involves monitoring people, so see the privacy section below.
Retail and warehousing
- Shelf monitoring: identifying gaps and misplaced products from shelf photos so staff can restock.
- Inbound checks: counting cartons, reading labels and spotting visible damage at the receiving dock.
- Barcode and label reading at angles and distances where handheld scanners are slow.
Logistics and field services
- Proof of delivery: checking that a delivery photo actually shows a parcel at a door.
- Meter and gauge reading from photos taken by field staff, reducing transcription errors.
- Vehicle damage assessment: comparing photos at pick-up and return.
Agriculture and food
- Grading produce by size, colour and visible blemishes.
- Crop health monitoring from drone or field images, flagging areas for an agronomist to inspect.
Construction and property
- Progress tracking by comparing site photos over time.
- Inspection support: highlighting possible cracks or corrosion in images for an engineer to review.
Office and finance
Not every computer vision application involves cameras on a factory floor. Reading invoices, receipts, identity documents and forms is computer vision too, combined with language processing.
What a successful project needs
Consistent image capture
Lighting, camera angle, distance and background affect accuracy more than most people expect. A fixed camera with controlled lighting over a conveyor is much easier than handheld phone photos in variable daylight. Improving capture conditions is often the cheapest way to improve results.
Labelled examples
Custom models learn from images that people have labelled: this one is good, this one has a scratch here. You need examples of every condition you want detected, including rare defects. If a defect appears only a few times a year, collecting examples is the hard part. Anomaly detection, which learns what "normal" looks like and flags departures from it, can help in that situation.
A decision about where processing runs
Edge processing runs the model on a device next to the camera, giving fast responses and keeping images on site, which suits production lines. Cloud processing is simpler to manage and scale but adds network delay and sends images off-site.
Integration with the process
A detection is only useful if it triggers an action: rejecting a part, creating a restock task, alerting a supervisor. Plan the connection to existing systems from the start.
Limits and risks
- Accuracy depends on conditions. A model that performs well in testing may degrade when lighting changes, a camera is knocked, or a new product variant appears. Monitor performance continuously.
- Errors in both directions. Missing a real defect and rejecting good items both cost money. Agree which matters more and tune for it.
- Privacy when people are in view. Monitoring workplaces or public areas raises legal and ethical questions, and laws vary by country. Use the minimum needed: detect a helmet without identifying the person, blur faces, and avoid storing footage unnecessarily. Facial recognition in particular is heavily regulated or restricted in some places and carries serious accuracy and bias concerns; avoid it unless there is a clear, lawful and proportionate need.
- Bias. Models trained on unrepresentative images perform worse on what they rarely saw, whether that is a product variant or a group of people.
- Hidden costs. Cameras, lighting, edge devices, labelling time and ongoing retraining all add up.
Running a sensible pilot
- Choose one narrow, high-volume visual check with a clear cost of errors.
- Fix the camera position and lighting before collecting data.
- Collect and label a representative image set, including awkward cases.
- Run the model alongside human inspectors and compare decisions.
- Only automate actions once agreement is reliably good, and keep humans reviewing uncertain cases.
Our AI and machine learning development team builds computer vision pilots, and our cloud solutions team can design the processing and storage behind them.
Key takeaways
- The best computer vision use cases are narrow, repetitive visual checks in controlled conditions.
- Image capture quality and labelled examples matter as much as the model.
- Plan the action each detection triggers and monitor accuracy over time.
- Treat any monitoring of people with care: minimise, anonymise and check local law.