VISIONROUTE

Glossary

Image recognition, defined precisely

The terms used across VisionRoute’s pages and reports, defined the way we actually use them — in product language first, engineering language where it earns its place. Each definition has a stable anchor, so you can link to a single term. No marketing inside the definitions.

Core concepts

Image recognition
Software determining what an image contains or what is true about it — which objects, text, categories, attributes, or rule violations are present. The everyday product term for the capability.
Computer vision
The engineering field covering how machines extract information from images and video. Image recognition tasks inside products are a subset of it.
Vision stack
The complete set of components that turns images into product decisions: the model or API, prompts or training data, preprocessing, output schema, confidence thresholds, review path, hosting, and monitoring. Products ship stacks, not models.
Visual use case
A recognition problem stated in product terms: the visual input, the decision to make about it, the output format, the user who acts on it, and the workflow it sits in.
Multimodal AI
AI systems that process more than one kind of input — typically images and text together. Vision-language models are the multimodal family most relevant to image recognition in products.

Approaches

Vision-language model (VLM)
A model that accepts images and text instructions together and answers in text or structured output. Enables recognition tasks to be specified in language, without collecting training data first.
Open-vocabulary detection
Locating objects in an image based on a text description of what to find, rather than a fixed list of trained classes. Useful when the classes change or were never trained.
Specialist model / API
A model or hosted service built for one recognition task — text extraction, face detection, logo spotting, safety screening. Strong on its task, fixed in scope and output format.
Custom-trained model
A model trained or fine-tuned on labeled examples from your own domain. Typically wins on unit cost, latency, and consistency once the task is stable and volume is high — at the price of data collection and MLOps.
Fine-tuning
Continuing the training of an existing model on your own examples so it fits your task better than prompting alone. Sits between prompting and training from scratch in cost and commitment.
Zero-shot
Using a model on a task it was never explicitly trained for, relying on instructions and general capability instead of task-specific examples.
OCR (optical character recognition)
Extracting machine-readable text from images or scanned documents. The mature, specialist end of image recognition; often one component inside a larger stack.

Evaluation

Representative data
Test images that mirror what production will actually see — common cases at realistic frequency, rare but costly cases, negatives, and real-world quality variation. The opposite of a demo set.
Expected answers (ground truth)
The correct output for each test input, defined by the product owner before testing. Every score is a comparison against these; without them, evaluation is opinion.
Acceptance criteria
The pre-agreed thresholds that decide pass or fail — e.g. minimum precision per rule, maximum tail latency, maximum cost per thousand items. Agreed before results exist, so results cannot renegotiate them.
Precision & recall
Precision: of the items the system flagged, the share that truly deserved flagging. Recall: of the items that deserved flagging, the share the system caught. Most recognition decisions trade one against the other.
Confidence threshold
The cutoff below which a system’s answer is not acted on automatically. Tuning it trades automation rate against error rate — and defines what goes to human review.
Abstention
A system declining to answer because confidence is too low, routing the case to a person instead. In production designs, a first-class outcome — not a failure.
Slice analysis
Reading results by meaningful subgroup — image quality, source, category, failure type — instead of one average. Production incidents almost always live in a slice the average concealed.
Regression suite
A versioned set of test inputs, expected answers, and scoring rules that can be rerun whenever models, prompts, or the product change — turning a one-time evaluation into ongoing protection.
Drift
Change over time that invalidates earlier results: providers update models behind stable names, your image distribution shifts, or the product’s definition of correct evolves. The reason evidence has a shelf life.
Hallucination
A model confidently reporting something that is not in the image — an invented field value, an object that is not there. Mitigated by structured outputs, confidence thresholds, and abstention, and measured by counting it as an error class in evaluation.

Operations

Inference
Running a trained model on new inputs in production, as opposed to training it. Inference cost and latency are per-decision operating numbers.
Structured output
Constraining a model to answer in a machine-readable schema (typically JSON) instead of prose — so the product can act on the answer without parsing free text.
Latency percentile (p50/p95)
Latency reported at the median (p50) and the slow tail (p95/p99) rather than the average. User experience and queue behavior are set by the tail.
Human-in-the-loop
A workflow design where people review a defined subset of automated decisions — typically low-confidence cases, appeals, and audits — so the system’s error rate is bounded by design rather than hoped for.

See the terms in use

In short

A vision stack is the complete set of components that turns images into product decisions — and the vocabulary above is how you specify, test, and operate one without hand-waving.

Applied in practice: an example free report · your own free App Scan