This article is dedicated to the development of a software system that performs person identification based on facial images using modern computer vision methods while addressing requirements for accuracy, speed, and adaptability. The following research is conducted in this work: analysis of contemporary biometric identification systems to determine their advantages and limitations; investigation of computer vision algorithms applied for face recognition and the justification of their selection; definition of system functional requirements; design of the software solution architecture and the database structure; implementation of the system's key modules according to the defined requirements; system testing and performance evaluation. A fully functional software complex has been created, encompassing everything from the video capture module to person detection and recognition. A crucial feature is the integration of an anti-spoofing module with two verification levels-neural networkbased and behavioral-which allows for reliable detection of fraudulent attempts. The developed software is suitable for practical implementation in controlled local conditions, namely at enterprises that require automated access control in real time. Moreover, thanks to its autonomy and flexible architecture, the system can be deployed even on robotic platforms for real-time access monitoring. This research also opens up prospects for further development in the areas of additional biometrics and improving verification algorithms.
As the days pass, we feel the increasing importance of security more acutely, both on a physical level and in the digital space. Intelligent access control systems are becoming particularly relevant. They offer a high-quality alternative to simply recognizing people by a password or conventional keys; instead, they use biometric data, which makes the system even more reliable, increases user confidence, and significantly simplifies life through the automated management of access to various facilities and data.
Thus, one of the most promising areas in this field is the use of computer vision for face identification. This approach ensures high speed, efficiency, and convenience in the identity verification process, and can be integrated into various systems—from corporate offices and schools to industrial sites. Although this technology is developing quite actively, certain obstacles remain on the path to its widespread implementation and use. Among these challenges are ensuring operational accuracy in various conditions, increasing the system’s security level against unauthorized access attempts, and the ability to adapt and update user data. That is why the development of a system that not only recognizes a person but also verifies their uniqueness and authenticity is a relevant and practically significant task.
The research task is to create a software system capable of identifying individuals by facial
image with a given accuracy and stability of operation. The system must process the input image,
extract a face from it, compare it with existing standards in the database and return the
identification result. In addition, it must ensure the storage of results in the database, allow
management of the composition of standards and interact with the user through the interface.
Thus, the creation of an effective personal identification system based on facial image analysis is a
complex interdisciplinary task that requires the involvement of modern computer vision methods,
data processing algorithms and software design principles. A comprehensive solution to this
problem will contribute to increasing the reliability of automated security systems and opens up
prospects for further research in the field of biometric authentication [
Thus, the creation of an effective facial recognition system based on facial image analysis is a complex interdisciplinary task that requires the involvement of modern computer vision methods, data processing algorithms, and software design principles. A comprehensive solution to this problem will contribute to increasing the reliability of automated security systems and opens up prospects for further research in the field of biometric authentication.
Access control is perhaps the most important element of creating secure conditions in information
and physical systems. As the speed of technological development is rapidly increasing, the need for
more reliable and convenient ways to identify individuals who have the right to access certain
resources is also increasing. Today, we know such traditional methods as: passwords, physical keys
and their modifications, PIN codes. However, they have quite significant drawbacks—they can be
forgotten, lost, transferred or forged. Given this, the optimal solution for ensuring confidentiality
has become biometric systems based on the unique characteristics of a person. Biometric data are
unique measurable physiological or behavioral characteristics of a person that can be used for
automated face recognition [
Biometric systems, unlike other well-known authentication methods, directly use the
characteristics of the person, which cannot be easily lost, forgotten or transferred to others [
The modern market is rich in various identification methods, for the implementation of which developers use their own set of technological tools. Let’s consider the main ones in more detail, taking into account their properties:
This method is used in smartphones, ATMs, corporate systems (for example, scanners in Dell offices). In this case, optical or capacitive scanners are used that read the unique pattern of papillary lines on the skin of the fingers. Based on this, certain digital templates are formed. This method is famous for its speed of information processing and is affordable in terms of the cost of equipment (scanners), but the presence of damage or other factors that affect the display of a fingerprint can reduce efficiency. allows you to create a detailed template of it. The disadvantages of this method include the need for special equipment, which causes difficulties in integration into real systems, although it is distinguished by high accuracy.
By palm geometry or vein pattern
Already used in medical institutions in Japan to access confidential data. These identification methods using specialized equipment can recognize a person by analyzing the physical parameters of the hand, its three-dimensional structure or special vein patterns under the skin. Unlike fingerprint identification, this method is more resistant to external damage and has an acceptable level of accuracy, but scaling to a large number of people may have certain limitations.
Voice
A behavioral feature used in banking call centers (e.g. Barclays Voice ID). This identification technology uses digital processing to analyze the acoustic features of the voice: intonation, frequency, timbre. Work can be affected by changes in a person’s voice due to health conditions or other physical conditions. The method is also sensitive to external noise, which negatively affects the quality of work.
By face
Used in airports, smartphones (Face ID technology in iPhone), video surveillance systems.
Systems based on this method of user recognition work by processing visual information from
images or videos. This allows you to create a vector description of the main features of a person’s
face. Modern computer vision and deep learning algorithms provide high accuracy even under
typically unfavorable conditions, such as poor lighting, partial face coverage, and head movements
[
Among the listed methods, one of the most popular and in practice the most convenient to use
is the facial recognition method, since it does not require direct contact, processes information
quickly, and the cameras necessary for its implementation are already equipped in most
environments [
Biometric identification systems for access control can be implemented both in physical objects
(for example, enterprises, educational institutions or airports) and in digital environments, such as
online banking, corporate platforms or mobile applications. In particular, facial recognition
technology, which is widely used by modern manufacturers to unlock devices or applications, is an
example of its effectiveness and high level of user trust [
Biometric systems have a number of advantages, including: high uniqueness of features,
difficulty of their forgery and automated identification process. Due to this, they are mainly chosen
for use in the security sector [
The use of biometric data for access control is one of the most promising areas in the
development of modern security systems. Despite the technical and ethical challenges, improving
algorithms, introducing new standards and integrating with IoT create conditions for the
implementation of continuous and adaptive authentication systems that can meet the requirements
of the future [
To implement the system, facial recognition was chosen as the main method of identifying individuals, since this approach does not require physical contact and is easily integrated into existing infrastructure, such as webcams and surveillance cameras. In addition, it provides high data processing speed and achieves a harmonious combination of high security and ease of use.
The face recognition system belongs to the category of biometric systems that use the analysis of a person’s appearance and characteristics for identification or verification. From a technical perspective, these systems utilize computer vision and machine learning algorithms.
The operating principle is as follows: first, they automatically locate the face in the image, then convert it into a numerical vector that describes its unique characteristics, and finally, compare this vector with previously stored data in the database.
As previously mentioned, the main advantage of this method is the absence of physical contact and the possibility of its implementation using ordinary surveillance cameras, which are already installed at the vast majority of facilities that might require user identification.
Modern facial recognition systems are based on a complex multi-layer architecture that
combines hardware and software algorithms for effective face identification. This allows processing
large amounts of data in real time, ensuring high accuracy and scalability [
Image capture module—responsible for receiving video or images from the camera in real time. At this stage, it is extremely important to obtain high-quality input data, as this directly affects the accuracy of recognition.
Face detection module—finds faces in the image among other objects. Methods such as Haar
Cascade Classifier or neural networks such as MTCNN (Multi-task Cascaded Convolutional Neural
Network) [
Normalization module—corrects the position, lighting, and scale of the image to prepare the face for processing. This stage allows you to reduce the impact of external factors on the quality of recognition.
Feature extraction module—builds a vector of facial features. For example, in the FaceNet model,
the face is converted into a 128-dimensional vector [
Matching module—calculates the distance between the vectors of the new image and those recorded in the database. If the distance is less than a certain threshold, the person is considered identified. In practice, Euclidean or cosine distance is usually used to calculate similarity.
Decision-making module—based on the comparison result, the system makes a decision to grant or deny access.
Systems can be implemented both in a local environment and on the basis of cloud platforms—
the choice depends on specific tasks, goals, security level and available resources. Local solutions
are usually chosen in cases where full control over the collected data is critically important [
Facial recognition systems for access control are usually supplemented with auxiliary modules.
These can be event logging systems that allow for activity auditing; an administrator interface for
responding to unauthorized login attempts, and notification modules for responding to situations
with suspicious activity. Some systems have an access filtering function, which is especially
important for large enterprises with a branched structure, as this allows them to be adapted to
security tasks [
Therefore, the architecture of such systems should be designed as flexible, modular solutions with the ability to scale. They should not only accurately identify the user, but also integrate computer vision algorithms with decision-making and data storage mechanisms to ensure accuracy and speed.
The current market for facial recognition systems offers a wide range of commercial solutions and open source projects that differ significantly in their functionality, level of accuracy, resource consumption, and areas of application. This section provides a comparative analysis of the most common systems, which will allow us to determine their advantages, disadvantages, and optimal areas of use.
First, let’s look at the most famous commercial systems: Apple Face ID
This technology from Apple is one of the most famous implementations of facial recognition in the commercial sector. It uses deep analysis technologies (TrueDepth camera) and 3D face modeling. Its great advantage is that all data processing is performed directly on the device, which minimizes the risks of leakage of confidential information, and also guarantees high resistance to spoofing. However, this technology is available only on a limited number of Apple devices and does not support integration into external systems.
Amazon Rekognition
This cloud-based facial recognition system from AWS offers APIs for integration with
thirdparty systems, which is a definite advantage. It also has a fairly wide functionality, which includes
facial recognition, emotion analysis, finding similar faces and comparing them with databases, and
supports integration with other AWS services, ensuring easy scalability. However, its operation
depends on a constant Internet connection, and the use of the technology raises discussions about
privacy and ethical aspects [
Clearview AI
A controversial technology, the implementation of which is based on a large database consisting of images illegally collected from social networks and other publicly available sources. According to reliable tests, it demonstrates high speed of work with large volumes of data, which, together with accuracy, makes it one of the favorites in the field of facial recognition. However, Clearview AI also has a significant disadvantage, which most experts note—violation of privacy through illegal accumulation of data and high risks of uncontrolled use.
At the same level as commercial systems, there are also open-source analogues that allow you to independently create custom systems without relying on third-party services:
OpenCV (cv2.face module)
The open-source OpenCV library includes the cv2.face module, which offers basic algorithms such as: Eigenfaces, Fisherfaces, and LBPH. These solutions are suitable for local projects with low accuracy requirements, as they are easy to use and do not require significant computing resources. However, compared to modern neural networks, the accuracy of such methods is limited.
Dlib + ResNet
This technology uses a 128-dimensional vector representation of faces for recognition. Dlib in combination with ResNet provides powerful recognition with high accuracy, if the data is properly prepared. The system performs well in situations where high accuracy is required, but requires significant efforts for high-quality training and computing power for optimal operation.
FaceNet / DeepFace
These open-source deep neural models—FaceNet and DeepFace, demonstrate excellent results in tests—accuracy of over 99% on standard datasets. Such a connection allows you to build quite complex recognition systems without the need for cloud environments. They are actively used in scientific research and commercial products. And although their main advantages are high accuracy and scalability, their implementation requires significant computing resources, which can complicate integration into simple local systems.
After conducting research and comparing all the analyzed models, we can conclude that it is open-source solutions that make it possible to develop an independent and flexible system with a level of accuracy sufficient for practical application.
In general, the ultimate technical goal of this work is to obtain a software solution for a local facial recognition system, which in turn creates security conditions. This system is built in a monolithic architectural style. This means that the entire application is created as a single technological system. All components are located on a single device, without interacting using external APIs or cloud environments. This implementation decision was made to increase the security and autonomy of the system.
It should also be noted that the system is based on a modular approach. The system architecture involves dividing it into logical components: capturing a video stream from a camera; detecting and highlighting faces in the frame; comparing embeddings with reference ones in the database; maintaining an event log (logging); administering users and access rights.
And although physically all these components are implemented as a single software, logical
division allows us to provide the modularity we are talking about. That is, all the mentioned
components have their own tasks, but they are all integrated into a single local program and work
harmoniously. In turn, this is done so that in the future there is an opportunity to scale the system,
expand its functionality without any cardinal changes and global restructuring of the architecture.
That is, we can conclude that the system is built mainly in a procedural style with a clear division
into functions (it uses a modular approach, when the main tasks are allocated into functional
blocks) [
No less important is the ability to scale. Although the system is local, the possibility of
expansion still remains: for example, you can add more users, increase the size of the database or
connect additional modules (emotion analysis, voice recognition or other biometric data), without
completely rebuilding the system core [
In order to prevent data leakage, the system is designed so that biometric features, namely
embeddings, which carry information about unique features, are stored exclusively in a local
database with strict access control [
Thus, taking into account all these factors, we obtain a system whose architecture is capable of
providing high data processing speed [
The development of modern facial recognition systems is a complex engineering task, therefore it requires the integration of various technologies and algorithms, which ensures high accuracy, speed and reliability. The correct organization of the components of such a system is the key to its successful functioning and further improvement. To achieve this result, it was decided to apply a modular approach to the architecture of the facial recognition system.
To ensure high flexibility, scalability and ease of system support, we will divide the system into modules, each of which is responsible for specific system functions. Their interaction creates the integrity and coherence of the entire mechanism.
This approach allows you to isolate the logic of different subsystems. This simplifies testing, makes it easier to detect and fix errors, and also provides the ability to expand functionality without a radical restructuring of the architecture.
To implement the functionality of a security system with automatic recognition of people, the following main modules were selected: video capture module; face processing module; face recognition module; database work and logging; anti-spoofing; administration.
It is thanks to them that we can provide full security guarantee functionality, which includes
face authentication itself, administration of this process and interaction, which directly allows for
dynamic communication with the system [
The video capture module is considered the first point of contact with the outside world. It is
responsible for stable and continuous receipt of a video stream in real time from a connected
camera device. In addition to basic frame acquisition, the module can also include pre-processing to
improve the quality of the input signal. It is able to adjust the exposure parameters, white balance
or apply simple noise reduction filters. To ensure compatibility with different types of cameras
(local, network, USB), the OpenCV library is used, namely its Python wrapper cv2, which provides
a universal access interface [
After that, we install the face processing module. Its task is to detect the presence of faces in the
received frames and prepare information for further recognition. To perform this task, the face
recognition library is used. It effectively uses the HOG algorithm implemented in the dlib library
[
Next, the found faces are determined using coordinates (bounding boxes), which allows them to
be accurately localized in the frame. After that, using the OpenCV library (cv2), each found face is
cropped according to the specified coordinates and scaled to the standard size required for further
processing. The main actions that occur in this module also include converting the color format
from BGR to RGB, which is necessary for face_recognition [
Another important function is face alignment based on keypoint detection using the same face_recognition. This allows you to reduce the impact of pose variations (tilts, turns) and significantly increase the accuracy of recognition. Then only the final stage is the formation of the famous embedding, which describes unique facial features and is the basis for further recognition.
The next module is the face recognition module. It is the “brain” of the system for recognizing
people. Here, the numerical face print (embedding) is compared with those that the system already
has. For this comparison, special tools are used that can determine how similar two values are.
Thus, the fact of coincidence between the embeddings is established and the proximity distance
between the vectors is calculated, which allows you to increase the accuracy and control over the
recognition process [
After that, the module gives the desired result or reports an unknown person. If there are
several faces in the frame, the system tries to focus on the main one (for example, the one in the
middle or the largest) [
The database module is used to manage the system data. Its work includes storing users’ biometric data (face embeddings), their personal information, and event logs.
Interaction with various types of relational databases is implemented through the PyODBC
library, which provides a universal interface via the ODBC standard [
The mandatory use of parameterized SQL queries is a key element for protecting against SQL injections and ensuring data integrity. In the future, the module can be extended to support various backup and recovery strategies.
The level of system security is extremely important in this work. This is provided by the
antispoofing module. Its main goal is to detect attempts to deceive the system by using fake facial
images [
At the moment, the integration of the Mediapipe library is planned. It provides tools for
analyzing key facial points and tracking barely noticeable micromovements, which is characteristic
of a living face [
Thus, combining a variety of anti-spoofing methods within this module will allow you to create a multi-level system of protection against unauthorized access.
The last is the administration module. It provides the administrative institution with a
convenient and understandable interface for performing key tasks of system management. This is
managing user accounts, and viewing system logs [
Here, the modular structure allows you to independently develop and update the administration interface without making changes to the basic logic of recognition and data processing, which is logical when building the system.
To ensure the coordinated, holistic operation of the system as a single complex, it is necessary to outline a clear interaction between individual modules. Each of them, performing its specific local task, in any case must interact with other modules through certain standardized interfaces. This ensures the independence of the modules, but at the same time guarantees the consistency of data processing. Let us consider in more detail the process of interaction of the modules within the framework of this study.
After starting the program, the system starts with the initialization of the image capture module. It establishes a stable connection with the camera and launches a continuous stream of frames in real time. This module also easily adapts everything to the use of different types of video cameras by changing the appropriate settings or drivers.
The captured frames are transferred to the second module—face processing. Here, the received information is analyzed for the presence of faces and this information is edited. Using the appropriate algorithms, the module determines the exact coordinates of each detected face in the frame, forms bounding boxes. Then the module crops the found face, scales it, and normalizes it.
The resulting cut-out face is processed in the anti-spoofing module. There, an analysis is made
as to whether a person is alive in the frame or not, and if it turns out to be real, the data is
transferred back to the processing module [
In the face recognition module, the face recognition itself is actually performed. Its data is compared with the existing ones (the process is described in more detail in the module description). That is, in parallel, we perform a query to the database using the database work module and “extract” the necessary records from there. It also enters logging information into the log events. Hence, this module serves as the main hub for both storing biometric data and logging.
Separately, we have interaction with the administration module. The results of the work
performed are transferred there [
So, having analyzed how the modules of this system interact, we can draw several conclusions. The key principles underlying this interaction are sequential data processing. This means that each module receives and processes information from the previous one, ensuring a controlled flow. In addition, the minimum level of communication between them, carried out through the transfer of standardized data, increases the system’s resistance to changes in the internal implementation of individual components. Therefore, thanks to these principles, the developed architecture is flexible, reliable and ready for further development.
The full functioning of a security system with automatic face recognition requires not only a direct, high-precision identification mechanism, but also a reliable infrastructure for storing and managing all important information. That is why an integral part of the architecture is the database subsystem, which ensures the preservation of information about registered users, administrative department data, as well as detailed logs of all key actions and events occurring in the system. Such a repository is the foundation for guaranteeing stability, security, auditability and further scalability of the system.
A relational database is used to store data, which acts as a central repository for biometric and auxiliary information. The main tasks of the database are: storing user face embeddings; storing information about system users (identifiers, names, additional attributes);maintaining event logs.
The selected database structure is based on the relational model. This is done to ensure the integrity and relationships between tables. This solution also allows you to build complex queries and easily scale the system when expanding the functionality. In addition, using the relational model allows you to organize data in a structured way in the form of tables with clearly defined fields and data types. This ensures ease of navigation, efficiency of query execution and flexibility of information processing, which allows you to easily implement connections between different entities and expand the logic of the system in the future.
For reasons of autonomy and increased security, it was decided to use a local MS SQL Server database. All data is stored locally on the device, eliminating the risks associated with transmitting biometric information over the network or using cloud storage.
For a visual representation of the organization of data in the database and a better understanding of its logic, let’s consider the description of the main tables and their fields.
Let’s start with the Position table, which stores information about employee job roles and allows you to keep track of structural positions.
The Guard table is responsible for representing the administrative level of the system. It stores information about the guards who manage access control to the system.
Next we have the table directly with users—(Table 3 Users). It contains data belonging to the system users, its visitors, and employees.
Here, the PositionID field is a foreign key that references the PositionID field of the Position table.
The Embedding table (Table 4) stores the embeddings, key biometric data of individuals, necessary for identification. This table also references the UserID field in the Users table using the UserID field.
UserID Surname Name Paternal Age PositionID DepartmentID PhoneNum PhotoPath Status CreatedAt
INT (PK) NVARCHAR NVARCHAR NVARCHAR INT INT (FK) INT (FK) NVARCHAR NVARCHAR NVARCHAR DATETIME Field size 100 100 100 13 255 50
Embed ID UserID Vector CreatedAt
INT (PK) INT (FK) VARBINARY DATETIME
MAX 100
NOT NULL NOT NULL NOT NULL NOT NULL NOT NULL NOT NULL NOT NULL NOT NULL NOT NULL
NOT NULL NOT NULL NOT NULL
Also, there is a table EntryLogs (Table 5) in the database. It is designed to keep a log of all access events, tracking each login attempt.
The table is referenced by the UserID field to the UserID field in the Users table, and the GuardID field is a foreign key that references the GuardID field in the Guard table.
The last table to describe is the Alerts table (Table 6). It contains information about alarming events. The table helps the administration to quickly respond to dangerous situations and keep track of responses.
AlertID LogID SentTime AdminResponse
INT (PK) INT (FK) DATETIME NVARCHAR
50
NOT NULL NOT NULL
Let’s present the main relationships in the database in a table, thereby demonstrating the relationship between the system entities:
To maintain data integrity and logical coherence, the database structure is built using clearly defined relationships between tables. This allows the system to efficiently combine information from different sources and quickly execute complex queries to obtain the necessary data.
The functioning of the automatic face identification system involves processing significant amounts of information in real time. A constant stream of images coming from surveillance cameras generates thousands of data reading, comparison and writing operations every second. Under such conditions, classical approaches to storing and accessing information are not effective enough, which necessitates the use of specialized data management methods focused on speed, scalability and load resistance.
In systems that implement computer vision, the main storage objects are feature vectors
(embeddings) obtained from neural networks during facial image processing [
One of the key areas of optimization is the use of a cache for short-term storage of processing results, which allows minimizing the number of repeated requests to the database and speeding up the system in conditions of a large stream of video data. Storing intermediate results in random access memory allows avoiding duplication of calculations and reducing the number of disk operations, which is especially important when processing streams in real time.
Several types of caching can be implemented in the developed system: results of the last successful user identifications. feature vectors for frames coming from the same camera during a short time interval. pre-normalized images ready for comparison.
For this, it is advisable to use high-speed in-memory storages, such as Redis or Memcached, which provide access to data with a delay of less than 1 ms. Redis has the advantage of supporting various data structures (dictionaries, lists, sets) and TTL (Time-to-Live) mechanisms, which allows you to automatically clean up outdated records and maintain the relevance of information. In the system, Redis can also act as a buffer between modules—for example, between image capture and recognition components, ensuring asynchronous data exchange.
To maintain the security of cached data, record lifetime control is implemented (30–60 s), after which they are automatically deleted, which makes it impossible to accumulate personal information. This approach guarantees both increased performance and compliance with confidentiality requirements.
Another important aspect of increasing the efficiency of the system is the indexing of vector data. For fast search of similar faces among a large number of records, traditional B-Tree or Hash indexes are inefficient, therefore specialized structures designed for vector search in multidimensional spaces are used. These include the algorithms FAISS (Facebook AI Similarity Search), Annoy (Approximate Nearest Neighbors Oh Yeah) and HNSW (Hierarchical Navigable Small World), which provide the search of the nearest neighbors (Nearest Neighbor Search) with sublogarithmic complexity.
The implementation of vector indexes allows you to search for identical or similar faces even with a large number of records (over 1 million vectors) with a response time of less than 50 ms. In the proposed architecture, it is advisable to use a database with support for vector types, such as PostgreSQL with the pgvector extension, or specialized repositories—\Milvus or Weaviate, which are focused on high-speed search of embedding vectors.
To ensure stable operation of the system with an increase in the number of connected cameras and users, the principle of distributed architecture is used, which allows you to effectively scale system components depending on the type of load. This approach provides the ability to simultaneously process dozens of video streams in real time without reducing performance or recognition accuracy.
The main modules of the system—image processing, anti-spoofing, identification, cache and database—can be implemented as separate microservices that interact via API or messaging system). This allows you to perform stream processing in parallel, as well as share resources between services depending on their functional role. For example, one node can process video from high-resolution cameras, while another—serves requests for identification or analytics. In such an architecture, scaling can be done in two ways: vertical scaling—increasing the resources of a single node (CPU, GPU, RAM) to process more complex recognition models; horizontal scaling—adding new nodes or containers that are automatically connected to the system using orchestration mechanisms.
The use of containerization allows you to quickly deploy new instances of services depending on the current load. For example, if the number of active cameras increases over a certain period of time, the system can automatically create additional copies of the identification service to maintain constant throughput.
In addition, load balancing between computing nodes ensures an even distribution of recognition requests, preventing overloading of individual servers. For this, both classic HTTPlevel solutions and specialized tools with dynamic routing control can be used.
An important element of the architecture is also monitoring the performance and status of components, which can be implemented through the Prometheus or Grafana systems. This allows you to track processing delays, resource usage, and the number of active threads in real time, which is critical for maintaining the stability of a highly loaded system.
Thanks to the implementation of these approaches, the system remains scalable, fault-tolerant and adaptive to load changes, maintaining minimal response time even during peak requests.
To store a large volume of access logs and events and at the same time ensure fast analytical queries, the principle of aggregated (summary) tables is used. Regular data aggregation by time intervals (hour, day, week) allows you to significantly reduce the execution time of administrator queries without losing the accuracy of statistics. Additional use of indexes on the UserID, CreatedAt, Status fields provides search optimization by the most commonly used criteria.
Implementation of the proposed approaches—caching, vector indexing and distributed data management—allows you to significantly increase the speed of the personal identification system without losing the accuracy of recognition. Thus, the system is transformed into a highly loaded information platform capable of scalability, fault tolerance and effective data flow management, hich fully meets the modern requirements for computer vision-based security system architectures.
In pattern recognition, information retrieval, and classification, precision [
TP = True Positive, i. e. when the actual value was “yes”, the model predicted “yes” (i. e. correct prediction = correctly worked) FP = False Positive, i. e. when the actual value was “no”, the model predicted “yes” (i. e. incorrect prediction = falsely worked)
TN = True Negative, i. e. when the actual value was “no”, the model predicted “no” (i. e. correct prediction)
FN = False Negative, i. e. when the actual value was “yes”, the model predicted “no” (i. e. incorrect prediction)
Accuracy is the proportion of all correctly classified data (positive and negative results). Mathematically, it is expressed as:
Accuracy= T P + TN
T P + F P + F N + TN
That is, it is the ratio of the number of correctly predicted cases to the total number of cases . An ideal model would have zero false positives and zero false negatives, and therefore the accuracy would be 1.0, or 100%.
Since it includes all four outputs from the error matrix (TP, FP, TN, FN), given a balanced dataset with the same number of examples in both classes, accuracy can serve as a general indicator of the quality of the model. For this reason, accuracy is often the default evaluation metric for typical models or those that do not perform special tasks.
The true positive rate (TPR), or the proportion of all actual positive outcomes that are correctly classified as positive, is also called completeness.
Mathematically, completeness is expressed as [
Recall = T P
T P + F N and describes how many objects the algorithm finds that are true anomalies—the percentage of true anomalies among all those that the system has identified as anomalies. The higher the accuracy, the fewer false positives, the higher the completeness, the fewer false positives. By changing the algorithm parameters, you can build a curve of the dependence of accuracy and completeness
Accuracy is the proportion of all results that the model classifies as positive and which are really such. Mathematically, it is expressed as follows:
Precision= T P
T P + F P and formally describes the confidence level of the classifier—the percentage of outliers found that are actually outliers.
The designed database structure demonstrates a thorough approach to ensuring both effective data organization and security. The described clearly defined relationships between tables guarantee data structure, where each record is logically linked to the corresponding entities, ensuring information integrity. Flexibility in forming complex queries to obtain the necessary data and ensuring system scalability is obtained. This allows adding new tables or fields without global restructuring of the schema. A well-thought-out relationship system is a guarantee of data integrity and correctness even with increasing load and number of users. In addition to the structure, special attention is paid to data security, which is critically important for a personal identification system. Transactional support guarantees the integrity of records during update and insert operations. A flexible audit system has also been implemented, which allows tracking all access attempts and administrative actions, which provides monitoring of suspicious activity and the possibility of post-disaster analysis. No less important is the protection of security guards’ accounts by storing hashed passwords. This measure protects the system from unauthorized login attempts. Thus, all these mechanisms together form a reliable and secure basis for the functioning of the personal identification system.
The results obtained confirmed the correctness of the chosen approach: the system correctly identifies users even on weak cameras and provides protection against spoofing. Declaration on Generative AI While preparing this work, the authors used the AI programs Grammarly Pro to correct text grammar and Strike Plagiarism to search for possible plagiarism. After using this tool, the authors reviewed and edited the content as needed and took full responsibility for the publication’s content.