Step 1. Online passport verification: what is required from the user
Automatic passport verification begins with a simple user action: uploading a photo or scan of the document to the system. Behind this simplicity lies a complex technological chain, where each element is critically important for successful verification. In the NeuroVision system, this stage includes detecting the document in the frame, checking the image quality, and verifying the set of uploaded pages according to the selected scenario.
The online passport verification process changed radically in 2024. The resource for checking against the list of invalid Russian passports, previously hosted on the website of the Russian Ministry of Internal Affairs (MVD), ceased operation. Organizations now use interagency electronic interaction systems and modern AI technologies to validate documents. This has led to stricter requirements for the quality of the data provided and the need to apply more advanced recognition algorithms.
For industrial-scale KYC, relying on single public forms for checking a document’s status does not meet compliance requirements: a reproducible check against several sources, the recording of results, and the possibility of a subsequent audit are needed. On the NeuroVision platform, passport verification is complemented by a database-check module (the KYC/AML loop), where a single process makes available a comprehensive check of an RF passport, checks against sanctions lists, the list of politically exposed persons (PEP), the list of terrorists, as well as additional checks that are often used in anti-fraud and credit scoring (FSSP, BKI, etc.).
What data is needed for automatic passport verification
For successful automatic passport verification, the system must obtain and process a strictly defined set of data. In the basic version of checking a document’s validity, the passport series and number are required, as well as the surname and first name of its owner. This minimal information makes it possible to perform initial validation through government systems.
Full-fledged passport verification for KYC (Know Your Customer) purposes requires an expanded set of data. The system must recognize and extract:
• The document series and number — a ten-digit code that is the unique identifier of the passport in government databases
• The owner’s full name — surname, first name and patronymic in exact accordance with the record in the document
• The date of birth — for cross-checking the age and correspondence with the customer’s other documents
• The subdivision code — a six-digit number allowing verification of the issuing authority
• The date of issue — critically important for determining the document’s validity period
• The place of birth — used for additional identification and compliance with AML requirements
For citizens of the Russian Federation, the automatic verification of such information is carried out by sending a request through the interagency electronic interaction system to the information resources of the Russian MVD. This means that each data element undergoes validation for compliance with government registries.
The NeuroVision IDP/AI-OCR system implements the processing of both printed and handwritten text, which is critical for older document formats and manually filled-in fields. The pipeline includes document classification and the extraction of data from documents of any type, with the subsequent transfer of the structured results to the organization’s internal systems.
For extended verification for financial organizations and telecom operators, the following may additionally be required:
• The registered address — extracted from the pages with registration marks
• Marital status — if the corresponding marks are present
• Information about previously issued passports — to build a complete history of the customer’s documents
Requirements for a passport photo for online verification
The quality of the passport image directly determines the success of its automatic verification. Modern systems require compliance with certain conditions: good light, no glare, a level image. These conditions are not arbitrary — each of them affects the ability of neural network algorithms to correctly process the document.
The main technical requirements for a passport photograph include:
| Category | Description |
|---|---|
| Lighting and exposure. | The image must be taken under even lighting without harsh shadows. Glare from lamination, watermarks and a complex background often cause recognition errors. Avoid direct light from a lamp or flash hitting the glossy surface of the document — this creates overexposed spots that make the text unreadable for the algorithms. |
| Full capture of the document. | The entire passport spread must fit into the frame, including all edges. Cropped document borders do not allow the system to correctly determine the document type and may lead to a refusal to process. At the same time, the passport must occupy at least 70% of the frame area to ensure sufficient detail. |
| Sharpness and resolution. | The minimum acceptable resolution is 300 DPI for scans or equivalent quality for photographs. Blurred images, where characters merge or become indistinct, cannot be processed automatically. Modern smartphones provide sufficient quality when the shooting conditions are met. |
| Shooting angle and geometry. | The photograph must be taken perpendicular to the surface of the document. The perspective distortions that arise when shooting at an angle complicate the work of the recognition algorithms. Although modern systems use modules for correcting the rotation angle and geometric distortions, straight-on shooting significantly improves recognition accuracy. |
| No extraneous objects. | The systems detect the fact that data is obscured by fingers and objects. Hands, stationery or other objects must not cover any parts of the document. It is also unacceptable to have other documents in the frame — the system must unambiguously determine exactly which document is subject to verification. |
Special requirements apply to the digital quality of the image. Modern verification systems detect the fact of shooting from a monitor, smartphone or tablet screen, and also detect photocopies of documents. This means that a passport photo taken from the screen of another device, or a scan of a photocopy, will be rejected by the system as potentially unreliable.
For passports with holographic security features, additional recommendations apply. Holograms must not create rainbow glare that interferes with reading the text. If necessary, it is worth taking several shots from different angles and choosing the option with the minimum distortion from the security features.
Compliance with these requirements ensures a recognition accuracy of 98-99%, which is critically important for the automatic processing of large flows of documents in banking, fintech and other industries where mass customer verification is required.
Step 2. Passport checking via OCR: how AI recognizes a passport from a photo
After the image is received, the IDP/AI-OCR loop is launched: the document is detected, classified, and converted from an image into a structured set of fields. In the NeuroVision system, the passport recognition process relies on a standard three-stage pipeline: finding the document’s edges, classification and data extraction, and the correction of rotations and distortions. The solution has been tested on more than 100 million documents and identifies documents with an accuracy of up to 99.97%. For industrial loads, speed and scalability are important: document recognition is performed in less than 1 second, which makes it possible to build a user verification scenario without waiting and without manual data entry.
Image preprocessing and locating the passport in the frame
The system’s initial task is to detect the passport in the photograph and prepare the image for recognition. The neural network analyzes the input image and determines the coordinates of the four corners of the document, even if the passport is photographed at an angle, partially obscured, or on a busy background.
Computer vision algorithms automatically correct perspective distortions, aligning the document in the plane. The system compensates for camera tilt, passport rotation, and the trapezoidal distortions that arise when shooting at an angle. After geometric correction, the image passes through filters for enhancing contrast and removing noise.
Intelligent algorithms adaptively adjust the brightness and contrast for each zone of the document separately. This is critically important with uneven lighting, when part of the passport is in shadow or overexposed by the flash. HDR processing technology restores the readability of the text even in overexposed areas.
Glare and shadow detectors automatically identify problem zones and apply local correction. Deblurring algorithms restore the sharpness of characters with slight camera defocus. The system also recognizes and compensates for the moiré patterns that arise when photographing laminated documents.
Recognition of the MRZ and passport fields using AI-OCR
After preprocessing, two-level recognition of the passport data is launched. The first level is the extraction of information from the machine-readable zone (MRZ), which contains the owner’s encoded main data. A specialized neural network is trained to recognize the characters of the OCR-B font used in the MRZ of all international-standard passports.
The algorithm sequentially reads the two lines of the MRZ, containing the document type, country code, the owner’s surname and first name, passport number, citizenship, date of birth, sex, and the document’s validity period. Built-in checksums make it possible to instantly verify the correctness of the recognition of each field.
In parallel, the second level works — the extraction of data from the visual zone of the passport. NeuroVision’s neural network models, trained on millions of document samples from 200+ countries, accurately determine the location of all information fields regardless of the language and design of the specific passport.
The system automatically recognizes text in Cyrillic, Latin, Arabic script and hieroglyphs. Multilingual OCR models correctly process diacritical marks, special characters, and national specifics in the spelling of names. Contextual analysis algorithms correct possible recognition errors of individual characters, relying on dictionaries of names and geographic names.
Semantic analysis technology automatically classifies the extracted information: it separates compound names, highlights patronymics, and correctly interprets various formats of dates and addresses. The system takes into account the specifics of passports from different countries — from American driver’s licenses with magnetic stripes to biometric EU passports with chips.
The final stage is the cross-checking of data between the MRZ and the visual zone. The information obtained from different sources is compared to identify discrepancies. Divergences between the machine-readable zone and the main passport fields may indicate a document forgery or an error in its production.
The entire recognition procedure takes less than a second, with the data extraction accuracy reaching 99.85% for quality images and remaining above 98% even for photographs of low resolution or taken in difficult conditions.
Step 3. Automatic passport checking against databases and rules
After the fields are recognized and structured, a check loop is launched: formal rules, anti-fraud signals, and cross-checking against data sources. On the NeuroVision platform, this stage is implemented through the AML-check module for individuals and entrepreneurs against 1,700+ databases, including global sources, with the possibility of regular continuous monitoring.
For passport KYC scenarios, targeted checks that are often required in compliance and anti-fraud are additionally applied: a comprehensive check of an RF passport, a check for debts with the FSSP, a check against sanctions lists, a check against the list of terrorists, a check against the register of politically exposed persons (PEP), as well as a check for being on the MVD’s wanted list.
Validation of the passport structure and checksums
The first line of defense is the mathematical verification of the correctness of the passport data. Modern passports contain a machine-readable zone (MRZ), where the information is encoded according to the ICAO Doc 9303 standard. Each line of the MRZ includes check digits calculated using a weighted modulo-10 algorithm. The system checks:
The checksums in the MRZ code. A single verification algorithm is applied to the document series and number, the date of birth, the expiration date, and the overall check digit. A mismatch of even one checksum indicates a recognition error or a forgery attempt.
The format and structure of the fields. The series of an RF passport must contain 4 digits, the number exactly 6 digits. The date of birth is checked for correctness (February 31 is impossible), and the owner’s age must correspond to the document type (an RF citizen’s passport is issued from the age of 14). The subdivision code consists of 6 digits in the format XXX-XXX, where the first three digits denote the region and the level of the subdivision.
The logical consistency of the data. The date of issue cannot be earlier than the date of birth plus 14 years (for a Russian passport) or later than the current date. The passport’s validity period must correspond to the age at which it was obtained: up to 20 years, from 20 to 45 years, and after 45 the passport becomes indefinite. The system also checks the correspondence of the subdivision code with the region of issue indicated in the passport.
Additionally, NeuroVision’s algorithms analyze the visual security features: the presence and correctness of watermarks, microtext, special fonts and copy-protection elements. The neural network, trained on millions of document samples, is able to detect anomalies in the placement of elements, violations in typographic printing, or the use of non-standard fonts.
Matching the passport against government and commercial registries
After successful structural validation, NeuroVision consults external data sources to confirm the document’s current status. This stage is critically important, since a formally correct passport may be invalid for many reasons.
A check against the MVD’s database of invalid passports happens in real time. The database contains more than 120 million records of passports that have lost their validity: lost, stolen, replaced upon a change of surname, expired, or seized by court order. The system checks the combination of the document series and number, receiving one of the statuses: valid, invalid with the reason indicated, or not found in the database.
Cross-checking against the FMS/GUVM MVD database makes it possible to confirm the fact that the passport was issued by a specific subdivision on the indicated date. The database contains information about all passports issued since 1997, including subdivision codes and the periods of their operation.
A check against the FSSP databases of debtors and enforcement proceedings reveals the presence of debts, restrictions on traveling abroad, and asset seizures. Although this information does not affect the validity of the passport itself, it is critical for financial organizations when deciding whether to grant a loan or open an account.
Sanctions and PEP lists include international (OFAC, UN, EU) and Russian lists of persons subject to restrictions. A check against the lists of politically exposed persons (PEP) is mandatory under the requirements of Federal Law 115-FZ on countering money laundering.
Commercial databases aggregate information from open sources: databases of court decisions, bankruptcy registries, lists of disqualified persons, affiliation databases. Modern systems are able to check a passport against more than 1,700 different sources, forming a comprehensive trustworthiness profile.
All queries to the databases happen in parallel, which makes it possible to obtain the full set of data in 1-3 seconds. If any database is unavailable, the system records this in the report, allowing a balanced decision to be made based on the available information.
Forming the final status of the online passport check
The final stage is the intelligent analysis of all the collected data and the formation of a single verdict on the document. NeuroVision uses scoring models, where each verification parameter receives a weighting coefficient depending on the criticality and reliability of the source.
The gradation of check results:
| Category | Description |
|---|---|
| Green status (successful verification) | Is assigned when all checks fully match: the document structure is correct, the passport is found in the MVD database as valid, and there are no critical risk factors. The probability of the document’s authenticity exceeds 99%. |
| Yellow status (additional check required) | Is issued when there are non-critical discrepancies: the passport is not found in the MVD database (possibly issued recently), there are minor debts, or technical inconsistencies in the image quality are detected. The system recommends a manual check by a specialist. |
| Red status (verification failed) | Means the detection of critical problems: the passport is listed as invalid, signs of forgery have been identified, checksum inconsistencies have been detected, or the person is on sanctions lists. The transaction is blocked automatically. |
The detail of the report includes not only the final status but also a detailed breakdown of each check: which databases were queried, what responses were received, the level of confidence in each parameter. The system records the time of the check, the versions of the databases used, and the validation algorithms applied.
Adaptive decision-making logic takes the context of the check into account. For opening a deposit, the requirements may be softer than for granting a loan. The system makes it possible to configure the threshold values and weighting coefficients for specific business processes, while complying with the mandatory regulatory requirements.
The check result is saved in a secure log with the possibility of an audit. This ensures compliance with regulators’ requirements and makes it possible, if necessary, to reconstruct the full picture of the decision made on a specific customer.
Step 4. Passport verification by photo and the user’s selfie
After the successful recognition of the passport and the check of the document against databases, the system moves on to the final stage — the biometric verification of identity. At this step, it is confirmed that the person who provided the document is indeed its lawful owner. In the NeuroVision face verification module, the comparison of the face in the document and in the selfie is performed in less than 0.1 seconds, and the identification accuracy reaches 99.97%.
Comparing the face with the photo in the passport (face verification)
Face verification technology extracts the biometric characteristics of the face from two sources: the photograph in the passport and the user’s current selfie. NeuroVision analyzes the unique parameters of the face — the distance between the eyes, the shape of the nose, the contour of the chin, the ratio of the parts of the face. The algorithm creates a mathematical model of the face (a biometric template) for each image.
NeuroVision’s algorithms are robust to changes in appearance: different lighting, shooting angle, the presence of glasses or a mask, age-related changes, makeup and hairstyle. The neural network is trained on millions of images and is able to correctly identify a person even if the photo in the passport was taken several years ago.
The matching process occurs through the calculation of the cosine distance between the vector representations of the faces. If the similarity score exceeds the established threshold (usually 0.7-0.8 depending on the required level of security), the system confirms the match. At the same time, the algorithm takes the image quality into account: blurred or darkened photos are automatically flagged for an additional check.
The liveness check and protection against forgery and deepfake
Liveness detection determines that a live person is in front of the camera, not a photograph, a video recording or a mask. There are two main approaches to checking for a live presence: active and passive.
In an active check, the system asks the user to perform certain actions: turn the head, blink, smile or say a random phrase. NeuroVision analyzes the naturalness of the movements, micro-expressions, and the reflexive reactions of the pupils to changes in lighting. The technology tracks the three-dimensional structure of the face and checks the correspondence of the movements to a person’s anatomical features.
The passive liveness check works imperceptibly for the user. The algorithm analyzes skin texture, glints in the eyes, facial micro-movements, and the pulsation of blood vessels. NeuroVision recognizes deception attempts: printed photographs give themselves away with moiré patterns and a lack of depth, device screens with characteristic glare and pixelation, and silicone masks with an unnatural texture and a lack of micro-movements.
To protect against deepfake, specialized detectors trained to identify the artifacts of generative models are used. The algorithms analyze the temporal consistency of frames, check the naturalness of facial expressions, and the synchronization of lip movements with speech. The detectors identify characteristic signs of synthetic images: inconsistency of lighting, artifacts at the borders of the face, unnatural transitions between frames.
How the system makes a decision on successful passport verification
The final decision on verification is made based on a comprehensive assessment of all the checks. The system uses a weighted risk assessment model, where each parameter is assigned a certain weight depending on its criticality.
The key factors for making a positive decision include: a match of the biometric data with a threshold above the established minimum, successful passing of the liveness check, the correspondence of the passport data to the results of the database check, and the absence of signs of image manipulation. Each factor is assessed on a reliability scale, and only when the cumulative threshold is reached does the system confirm successful verification.
When suspicious signs are detected, the system may request an additional check: a repeat selfie of improved quality, a video recording performing instructions, or an alternative method of confirming identity. All verification attempts are logged for subsequent analysis and improvement of the algorithms.
The verification result is formed as a structured report indicating the level of reliability of each check. The client’s business logic can configure the threshold values depending on the security requirements: for financial operations, the strictest parameters are set, and for registration in loyalty programs, softer ones.
Modern online passport verification is a multi-stage process in which artificial intelligence sequentially recognizes the document, matches its data against government registries, and confirms that a real person is in front of the camera, not a photo or a forgery. This approach eliminates the manual processing of applications, reduces verification time to a few seconds, and at the same time increases protection against fraud thanks to the checking of checksums, MRZ zones, and liveness tests.
Combining AI-OCR, biometric face comparison, and anti-fraud algorithms makes it possible for a business to automate customer identification without losing accuracy and reliability. Companies gain the ability to scale verification processes, reduce operating costs, and comply with regulators’ requirements, offering users a convenient and fast way to confirm their identity in the digital environment.
