Computer vision is a field of artificial intelligence that teaches programs to extract meaning from images and video: find a passport in a photo, read its fields, compare a face with a photograph, spot a person without a hard hat in a frame. The technology is already at work in banks, fintech, retail, and manufacturing — wherever these tasks need to be solved in seconds. Let’s look in detail at what computer vision is, how it works, what tasks it solves, and where it makes mistakes — with examples from document, face, and video verification.
What Computer Vision Is in Simple Terms
Computer vision (CV) is a field of artificial intelligence in which methods teach programs to extract information from an image or video about what it contains: which objects, where they are located, what is happening in the frame.
Source: Decree of the President of the Russian Federation No. 490 of 10 October 2019, clause 5(b) (version of 15 February 2024) — computer vision in the list of AI technologies
Image processing changes the picture itself — it removes noise, increases sharpness, rotates the frame. Computer vision draws a conclusion about the content of the picture — “this is a Russian passport, here is its number.” In a real system, processing comes first and analysis second.
Machine vision is the term traditionally used for industrial systems: a camera on a conveyor, defect inspection, fixed shooting conditions. Computer vision is a broader field: shooting on a phone, video streams, biometrics. There is no strict boundary between the terms; this is established usage.
The National Strategy for the Development of Artificial Intelligence (Decree of the President of the Russian Federation No. 490 of 10 October 2019, clause 5, version of 15 February 2024) names computer vision first in its list of artificial intelligence technologies — alongside natural language processing and speech recognition and synthesis. These concepts apply for the purposes of the Strategy itself — this is a definition from a strategic planning document.
For a computer, a photograph is a table of numbers. The same passport at a different angle, under different lighting, or with glare produces different numbers, even though the meaning has not changed. That is why, for a neural network, “seeing” means learning to ignore whatever does not change the meaning.
How Computer Vision Works: From Pixels to a Decision
An Image Is an Array of Numbers
A pixel is brightness across the three RGB channels, usually from 0 to 255. A Full HD frame is about 6 million numbers (1920 × 1080 × 3). Video is a sequence of frames, usually 25–30 per second.
Features: At First, Humans Designed Them
Until the 2010s, engineers manually described what exactly to look for in an image:
- brightness changes and contours;
- rectangular features resembling Haar basis functions — in the Viola–Jones face detector (2001): it ran in real time — 15 frames per second on a 2001 computer;
- histograms of oriented gradients, HOG (Dalal and Triggs, 2005) — for finding people in photos.
This approach broke down when the lighting or the angle changed.
Neural Networks Learn Features Themselves
A convolutional neural network (CNN) passes the image through layers of filters: the first layers respond to edges and blobs, the next ones to corners and textures, the deep ones to parts of objects: an eye, the corner of a document, a coat of arms.
The idea goes back to Hubel and Wiesel’s experiments with a cat’s visual cortex (1959), Fukushima’s neocognitron (1980), and backpropagation for reading postal codes by Yann LeCun and colleagues (1989). Since 2020, transformers (Vision Transformer) have worked alongside convolutions: the image is cut into 16 × 16 pixel patches, and the model evaluates the relationships between them.
Training on Labeled Examples
The model sees thousands or millions of images with correct answers — a box around a document, the text of a field, “the same person or different people” — and adjusts its weights until the error is small. Quality depends on the data no less than on the architecture: if a country’s documents are scarce in the dataset, the model makes mistakes on them more often.
The Answer Is a Probability, and the Decision Is a Threshold
The model does not say “yes” or “no” — it outputs a confidence (score). The decision threshold is chosen by the business: above it, the system decides on its own; below it, the case goes to an employee. The threshold determines the balance between two errors: letting an impostor through and rejecting a genuine user. How this pair is measured in practice (FAR and FRR — the false acceptance and false rejection rates) is covered in the article “Face Recognition System: How It Works, Its Accuracy, and Where It’s Used”.
What Tasks Computer Vision Solves
In practice, computer vision covers several typical tasks.
| Task | What the model answers | Example in customer verification and video analytics |
|---|---|---|
| Classification | “what is in the picture as a whole” | a Russian passport, a driver’s license, or another country’s ID card |
| Object detection | “what and where,” bounding boxes | where the document, the face, or a hard hat on a worker’s head is in the photo |
| Segmentation | which pixels belong to what | precise document boundaries, separating the face from the background |
| Text recognition (OCR) | the text and its fields | full name, number, date of birth, machine-readable zone (MRZ) |
| Key points (landmarks) | coordinates of characteristic points | eyes, nose, corners of the mouth for face alignment |
| Vector comparison (embeddings) | how similar two images are | a selfie and the passport photo — whether it is the same person |
| Motion analysis and tracking | how an object moves between frames | crossing a line, time in a queue |
| Image authenticity check | whether a live subject was captured | a photo of a screen, a mask, a deepfake, traces of document tampering |
Real products rarely solve just one such task — they usually combine several. Document recognition, for example, is classification, detection, OCR, and rule-based checks in a single pass.
A Brief History of Computer Vision: From MIT’s Summer Project to Neural Networks
The field is older than it seems: the first work began in the 1960s, and the breakthrough came in 2012.
| Year | Event | Why it matters |
|---|---|---|
| 1959 | D. Hubel and T. Wiesel describe neurons in a cat’s visual cortex that respond to edges of a specific orientation | the idea of a hierarchy of simple and complex detectors, later carried over into neural networks |
| 1963 | L. Roberts, MIT dissertation “Machine Perception of Three-Dimensional Solids” | one of the first works on recovering 3D shape from photos |
| 1966 | S. Papert, MIT AI Memo 100 “The Summer Vision Project” | a summer student project to “teach a computer to see”; the task turned out to be decades harder |
| 1980 | K. Fukushima, the neocognitron | a multilayer network robust to shifts — the prototype of convolutional networks |
| 1989 | Y. LeCun et al.: backpropagation for handwritten postal codes | the first practical convolutional network; by 1998, convolutional networks from the same work were already reading several million bank checks a day |
| 2001 | P. Viola and M. Jones, a cascade face detector | real-time face detection on an ordinary computer of the time |
| 2005 | N. Dalal and B. Triggs, HOG | reliable detection of people in photos |
| 2009 | The ImageNet dataset (Deng, Fei-Fei Li, et al.) | according to the ImageNet project, more than 14 million labeled images; since 2010 — the annual ILSVRC competition |
| 2012 | AlexNet (Krizhevsky, Sutskever, Hinton) | at ILSVRC-2012, one of the model’s variants achieved a top-5 error of 15.3% versus 26.2% for the runner-up; the start of the deep learning era |
| 2015 | ResNet (He et al.) — an ensemble with a top-5 error of 3.57% at ILSVRC 2015; FaceNet (Google) — face comparison via vectors | very deep networks; the embedding approach on which modern biometrics is built |
| 2020–2021 | Vision Transformer (Dosovitskiy et al.) | transformers come to vision |
The 2012 leap came from three things at once: large labeled datasets, graphics cards (GPUs), and deep networks. The same principle holds today: a model’s quality is determined by the data it was trained and tested on.
Computer Vision in Customer Verification (KYC, Know Your Customer)
Remote identity verification is built on computer vision: the customer sends a photo of a document and a selfie, and within seconds the system decides whether the document is genuine and whether it is the same person. More on this kind of verification in the article “KYC in 2025: How Customer Identification Works”. What follows uses examples from NeuroVision’s computer vision platform.
Documents: Recognize, Read, Verify
The model determines the country and type of the document, finds its fields, and reads them (OCR). It checks the machine-readable zone (MRZ) against the ICAO Doc 9303 standard: check digits and consistency with the visual zone. Separately, it looks for signs of forgery and tampering.
Document recognition, according to that page, works with more than 16,000 document types from 195+ countries and territories, with a field recognition accuracy of 99.85%, and takes about a second per document. How such systems detect forgeries is covered in the article “How KYC Systems Detect Forged Documents”.
Face: Matching a Selfie With the Document Photo
A different task is solved here — 1:1 comparison: the model builds face vectors from the selfie and from the document and compares them. This is a one-time match within a single check, not a database search. How it is measured with the FAR/FRR pair is covered in the article “Face Recognition System: How It Works, Its Accuracy, and Where It’s Used”. According to the product page, matching accuracy is 99.74%, and the algorithm achieved the best result among participants from Russia in the NIST FRVT (Face Recognition Vendor Test) test (March 2023).
Liveness: A Live Person in Front of the Camera
A separate task is to distinguish a live face from a photo on a screen, a printout, a mask, or a replayed video. Such attempts are called presentation attacks. A detailed breakdown is in the article “Liveness Detection: What It Is and How Liveness Checks Work”. According to the product pages, the accuracy of the liveness check is 99.9%.
Deepfakes and Video Stream Spoofing
Computer vision works on both sides here: generative models create fake faces and videos, while detectors look for synthesis artifacts and signs of a virtual camera. More in the articles “Deepfakes in KYC: Detecting Face Spoofing and Protecting Video Verification” and “How Video Streams Are Spoofed in KYC”.
The NeuroVision platform combines document recognition for 195+ countries and territories, face-to-document matching, and liveness in a single KYC session. Integration is via REST API and SDKs for Web, iOS, and Android; in the cloud or installed within the company’s own perimeter (on-premises).
Computer Vision on Camera Streams: Video Analytics
The same technology works with ordinary surveillance cameras: the model analyzes the frames itself, and the operator receives an event instead of hours of footage. This area is called AI-based video analytics. Typical examples:
- monitoring of personal protective equipment (PPE: hard hat, vest, goggles) and hazardous zones in manufacturing — more: PPE monitoring in manufacturing;
- counting visitors and queues in a store — anonymously, only the number of people in the frame — more: store visitor counting;
- perimeter and entry control and vehicle license plate recognition;
- finding the right fragment in the footage archive.
People counting and PPE monitoring do not recognize a specific person.
Automatically recognizing a person from a stored face template is a separate task — identification — and Russia’s biometrics law (Federal Law No. 572-FZ) regulates it separately from authentication. In organizations’ information systems, processing biometrics for identification purposes — including collecting and storing them — is prohibited, and the only exception to this prohibition is collecting biometrics for placement in the Unified Biometric System (EBS), Russia’s state biometric information system, in accordance with federal laws (Article 15(1)(1) of 572-FZ). Identification through the Unified Biometric System is possible in cases established by federal laws, acts of the Government, and regulatory legal acts adopted in accordance with them (Article 9(1) of 572-FZ).
Authentication is a separate regime: confirming that biometric data already provided matches the vectors stored in the Unified Biometric System, subject to the conditions of Article 16(1) and (3) of the law (Article 15(1)(2)). This is a “one-to-one” match, not a search for the right face among many.
There is also a third case — a boundary of the law’s scope, not a path the law permits: the biometrics law does not apply to identification and authentication if the biometric match is checked without automation, with the participation of an authorized official (Article 1(2)(2)). This wording should be read as “the law does not reach here,” not as “the law permits this” — and the personal data law continues to apply. Whether a private security guard is an “authorized official” within the meaning of this provision, the provision itself does not define — the question remains open.
Organizations are prohibited from keeping biometrics for identification in their own possession, outside the Unified Biometric System (Article 15(1)(1) of 572-FZ). And recognizing a person by their face without their written consent — outside the cases that the personal data law names separately (for example, operational-search activities or counter-terrorism, Article 11(2) of Federal Law No. 152-FZ “On Personal Data”) — violates Article 11(1) of the same law.
More on the Unified Biometric System itself in the article “Unified Biometric System (EBS)”.
Footnote: Paraphrase of the provisions as of 2 November 2026. 572-FZ — Article 1(2), Article 15(1) — version of 28 December 2024; 152-FZ — Article 11(1) version of 25 July 2011, Article 11(2) version of 23 May 2025. The legal assessment of a specific setup is for a lawyer to make.
NeuroVision video analytics connects to existing cameras and sends alerts about PPE violations, entries into zones, and queues.
Where Else Computer Vision Is Used
Computer vision is used far beyond customer verification:
- Industry — quality control on the conveyor (machine vision), defect detection, reading markings.
- Medicine — helping doctors analyze images: X-ray, CT. The decision remains with the doctor, and in Russia such programs are registered as medical devices.
- Transportation — recognition of license plates, road markings, and signs; driver assistance systems.
- Retail — visitor counting, shelf display monitoring, self-checkout.
- Agriculture — assessing crops from drone and satellite imagery.
- Smartphones — face unlock, photo processing, translating text from the camera.
More examples by industry are on the industry solutions pages.
Limitations: Where Computer Vision Makes Mistakes
Capture quality. Glare, blur, a cropped document edge, poor lighting — the main reason for repeated capture attempts. More in the article “Why Users Retake Documents”.
Training data and demographic differences. In its NISTIR 8280 report (December 2019), NIST showed that for many face recognition algorithms the error rate differs by sex, age, and origin; the magnitude depends on the algorithm — accuracy is worth testing on your own audience.
Adversarial examples and attacks. As early as 2013, Szegedy and co-authors showed that barely noticeable pixel changes can fool a neural network. In KYC, this role is played by forgeries, masks, video replay, and video stream injection — which is why liveness and anti-fraud are needed.
Drift. Cameras, document blanks, and fraudsters’ techniques change. The model needs to be periodically retrained and retested.
Personal data. A facial image used to establish a person’s identity is biometric personal data (Article 11 of 152-FZ). Hence the requirements for consent, storage, and the place of processing.
Computer vision produces a probabilistic assessment. In processes with legal consequences, disputed cases are reviewed by a person, while the system retains the frames and the reasons for the decision for review.
How to Evaluate a Computer Vision Solution Before Deployment
A practical checklist for anyone choosing a vendor:
- Task-specific metrics. For detection and OCR — precision and recall, the share of correctly read fields. For biometrics — FAR and FRR. For liveness — APCER and BPCER.
- Independent tests. For faces — the NIST FRTE program (until 18 August 2023 — FRVT, then split into FRTE and FATE). For liveness — testing under ISO/IEC 30107-3 in an accredited laboratory.
- A pilot on your own data. Your own documents, cameras, audience. Compare with your current manual process.
- Speed and load. Response time per check and behavior under peak traffic.
- Where the data is processed. A cloud in Russia or installation within your own perimeter (on-premises) — this determines the information security and 152-FZ requirements.
- Explainability and logging. Can you see why the system rejected: which field, which signal, which threshold.
More on vendor requirements in the article “KYC Provider Requirements”.
Glossary of Terms
| Term | What it means |
|---|---|
| Computer vision (CV) | a field of AI in which a program extracts information from an image or video about what it contains |
| Machine vision | a term for industrial vision systems with stable shooting conditions, for example a camera on a conveyor |
| Convolutional neural network (CNN) | a neural network that learns by itself to find features in an image through layers of filters |
| Embedding (feature vector) | a numerical representation of an image or a face, used to compare two objects |
| Detection | the task of finding an object in an image and indicating its bounding box |
| Segmentation | the task of determining exactly which pixels of an image belong to an object |
| OCR | text recognition in an image |
| MRZ | the machine-readable zone of a document — lines with check digits at the bottom of the page |
| Liveness / PAD | a check that a live person is in front of the camera (PAD — presentation attack detection) |
| FAR / FRR | a pair of errors in biometric comparison: false acceptance of an impostor (FAR) and false rejection of a genuine user (FRR) |
FAQ
Computer vision turns a frame from a camera or scanner into objects, text, and faces that can be compared and checked. Its quality depends on the architecture, the training data, and the chosen threshold — which is why a solution is worth testing on your own scenarios. In customer verification, computer vision covers the document, the face, and liveness; in video analytics — events in the frame. And recognizing a person by face without written consent or outside the Unified Biometric System is restricted by law separately.
