AI and computer vision glossary

Artificial intelligence applied to physical operations such as production lines, food and packaging plants, warehouses, quarries and worksites. Most often it means AI vision that inspects products or watches over a site, running on site and connected to the operation’s existing control and business systems.

Camera inspection where a trained AI model, rather than fixed rules, decides whether an item is good or faulty. It learns from example photos, so it handles natural variation in colour, position and lighting, and can be retrained when products change.

Automatic camera checks on products as they move along a line, looking for faults such as surface defects, missing parts, wrong labels or damaged seals. It can be rule-based machine vision, AI vision or both, and usually covers every item.

The field of AI that lets software interpret images and video: detecting, counting, classifying and measuring what a camera sees. Inspection on a production line and monitoring on a site are both applications of computer vision.

The industrial use of cameras for inspection, measurement and guidance. Traditionally it means purpose-made cameras, controlled lighting and rule-based image processing (measure this edge, check that colour), which excels at precise, repeatable measurement. AI vision brings trained models to machine vision for faults that vary, and can be added alongside an existing system.

Finding faults on products as they’re made: surface marks, cracks, dents, misalignment, missing parts, colour variation or contamination. AI defect detection learns what each fault looks like from photos of your own product.

Checking that the printed date or batch code on every pack is present, readable and correct for the product running, so mislabelled stock doesn’t ship. One of the most common checks on packaging lines.

Optical character recognition (OCR) reads printed text such as date, batch and lot codes. Optical character verification (OCV) then checks it against the code expected for the run.

Removing the previous run’s product, labels and packaging before the next run starts, then confirming the right materials are loaded. AI vision can check the first packs against the new product and keep an image record for the plant’s own sign-off.

A good item wrongly rejected by an inspection system. False rejects waste good product, so they are tracked as closely as missed faults. Training on real examples of good product, agreeing thresholds with the quality team and validating the model on production data before deployment keep them down.

A routine check where known-bad samples are run through an inspection system to confirm each fault is still detected and removed. Results and images are kept with the plant’s quality records.

Teaching an AI model a task by showing it labelled examples, such as photos of good product and of each fault that matters. The trained model is then tested on images it hasn’t seen before it’s trusted on a running line.

A vision model trained on images of your own products, line or site, rather than a generic model trained on internet photos. Most industrial work needs one, because generic models don’t know your products, packaging or faults.

The Industrial AI Vision Lab, or VLab, is Industrial AI’s model-training environment. Product-specific vision models are developed there, trained on a customer’s own products and defect types, and validated against their production footage before deployment.

Short for “You Only Look Once”: a widely used family of real-time object detection models that analyse a whole image in a single pass. Their speed makes them a common building block for vision systems that run on site.

Connecting an AI vision system with a plant’s existing control systems, reject equipment, operator displays and data platforms, so each inspection result drives action on the line and reaches the people who need it.

The equipment that removes faulty items from a line, such as an air jet, pusher, diverter or flap gate. In an AI vision setup, the vision system signals the line’s existing reject equipment through the plant’s control systems.

Programmable logic controller: the industrial computer that runs a machine or line, reading sensors and switching outputs like motors, valves and reject gates. A vision system passes its result to the line’s controls, and the controller decides what happens next.

Supervisory control and data acquisition: the software that monitors and records a whole plant across many machines and controllers. Vision results such as counts and reject reasons can be passed to it, so they sit alongside the plant’s other data.

Human-machine interface: the touchscreen or panel operators use to run and monitor a machine. Vision results, alarms and quality data can be presented on existing operator displays as part of plant integration.

Running AI on hardware within your facility instead of sending images to the cloud. Decisions are made locally, inspection has no dependence on external connectivity, and footage stays on your premises.

Running a trained model on new data to get an answer, as opposed to training it. In vision inspection, inference runs on every image captured, usually on edge hardware on site.

The principle that data is stored and processed under the laws of the place it was created. For New Zealand operators it often means camera footage and production data stay on site or in New Zealand. This is general information, not legal advice.

The first step in a vision project: a review of the products or site, cameras and lighting, and the faults or events that matter, with a recommendation on how AI applies and what it would need.

Using cameras and AI to count loader cycles into each truck at a quarry or aggregate site, with a timestamped image for each load. This supports weighbridge reconciliation without measuring weight or replacing the weighbridge.

Automatic number plate recognition: computer vision that reads vehicle number plates from camera footage. It’s used at site gates, depots and weighbridges to log vehicle movements automatically.

Computer vision that checks whether people are wearing the required protective equipment, such as hard hats and high-visibility vests, in defined areas of a site. It’s usually paired with exclusion-zone monitoring.

Software powered by an AI model that takes actions rather than only answering questions: reading and replying to email, updating systems, drafting quotes and following up. It works inside your tools, within limits you set.

An AI agent that answers phone calls in natural speech. It answers common questions, books jobs and takes messages, and every call is logged.

An AI model trained on large amounts of text that can read, write, summarise and reason in everyday language. Claude, the models behind ChatGPT, Gemini and Llama are examples. LLMs generate likely text rather than look facts up, so business uses pair them with your own information.

Retrieval-augmented generation: the AI first looks up the relevant parts of your own documents, prices or policies, then answers using them. It keeps an AI agent’s answers grounded in your business rather than in general training data.

Software that runs a multi-step business process by itself: an enquiry arrives, is sorted, a quote is drafted and sent, and a follow-up is scheduled. Modern automation mixes fixed rules with AI for the steps that need judgement.

An open-source AI agent framework that runs on your own devices and connects to channels like WhatsApp, Telegram, Slack and email. It’s used to run a private AI agent that handles tasks such as enquiries, documents and scheduling.

Anthropic’s AI coding agent. It reads and edits code, runs commands and works through engineering tasks from the command line, which lets teams build internal tools faster.

Model Context Protocol: an open standard, introduced by Anthropic, for connecting AI models to tools and data such as databases, email and business apps, without writing custom code for every connection.