Now, face recognition products on the market have achieved good detection results in customs clearance, finance, telecommunications, notarization and other scenes that need the same witness. Applications in transportation, public security, buildings, communities and other fields are also gradually developing.

At present, the face recognition system in the security system generally refers to the system based on the dynamic face detection, recognition, alarm and query in the surveillance video. The face recognition system mainly includes face image acquisition, preprocessing, feature extraction and matching recognition. The existing face recognition technology is mainly based on the face recognition of visible light image.

Due to its own basic technical conditions, and the performance optimization of specific scenes, all kinds of face recognition products will reflect different application effects and differences. Moyan is a face recognition product with high performance and reliability. The network terminal product is composed of face recognition terminal, humanoid gateway and client management system. Through the ID card information collection at the entrance and exit, real-time face capture and ID card comparison, users can pass quickly and efficiently. The evaluation from the perspective of Huaxia Zhixin’s “Moyan” recognition, face recognition efficiency, dark environment, multi person crossing, multi person recognition, safety test, etc., strive to objectively restore the performance of the product in practical application.

Test occasion and matching equipment

Huaxia Zhixin “Moyan”, Zhongan laboratory, tripod, camera, illuminance meter, handheld pan tilt, auxiliary props, PC, PS software, etc.

Performance evaluation of Huaxia Zhixin "Moyan" face recognition product

Unpacking and appearance

Unpacking

display

The package of the whole machine includes:

1. Moyan noumenon

2. Hexagonal base

3. Base gasket

4. Power supply

5. Fixing ring

In terms of hardware, Moyan has a classic structural design and appearance. It is equipped with an 8-inch high-definition display screen, and an advertising space is reserved at the bottom of this screen to play value-added ads. At the same time, Moyan supports various mainstream interfaces, as shown in the figure below:

Angle adjustment

Vertical adjustment

Level adjustment

The test results show that the upward movable range of the terminal is the largest, and the upward angle can vary from 0 ° to 80 °; The downward movable range of the product is the second, and it varies from 0 to 25 degrees; The horizontal movement of the product is relatively limited, and the effect is as shown in the figure, which can be changed in the range of 0-20 degrees.

Rapid deployment

This test mainly uses the stand-alone function of the terminal. After power on, the system will directly enter the face recognition page. At this time, you can enter the face recognition setting page by continuously clicking the top right of the screen for four times. At the bottom of the page, you can choose to enter the on-site face registration step.

The testers face the camera in front, take a picture for the subsequent face recognition test, then fill in the relevant information according to the system instructions, save and submit, and the system shows that the registration is successful.

The registration process of single machine mode is very simple. However, if you want to batch import, you can only select the text pre entry or batch import function. At present, the two functions are independent of each other. The text pre entry function can realize batch import of character information into excel, and the batch import function can realize batch import of pictures.

After completing the face registration step, you can enter the specific test stage

Test items

Limit recognition angle

Up

down

towards the left

towards the right

The GIF image record reflects the limit angle that the terminal can correctly recognize the face with the increasing recognition angle in all directions.

The final results show that the recognition limit angle of the terminal is about 50 ° in the vertical direction and 30 ° in the horizontal direction.

Face recognition efficiency

In order to test the recognition effect when the face is blocked, the testers simulate wearing masks, hats, sunglasses and scarves to carry out the recognition test.

When the tester wears a hat and sunglasses to cover the upper part of the face, the terminal will have a high recognition success rate as long as it detects the features of the person’s eyes. It should be noted that as long as the tester covers the lower part of the face, the terminal will show recognition failure. Only when the mouth is exposed can the terminal recognize normally.

Backlight and dark environment performance

The minimum light intensity required for recognition is about 10 lux, which is about equal to the natural light intensity at 6 pm.

Below this extreme value, we need to turn on the infrared detection light compensation strategy.

Infrared somatosensory

In the backlight scene, the natural light intensity at noon on a sunny day is about 7000 lux, which is used as the strongest light to test, and Moyan can also perform face recognition smoothly.

When used indoors, the illumination changes from tens to hundreds of lux, so Moyan can complete face recognition accurately and quickly.

Therefore, it can be seen that Moyan adapts well to different lighting conditions and can be applied to various indoor scenes.

Passing through the gate without feeling

Many people pass the gate

The first group of tests is to simulate the normal face clock in scene. The testers stop when passing the terminal, and turn their upper body to the terminal, so as to leave enough time for the terminal to recognize and retrieve.

Through the GIF diagram, we can see that Moyan can recognize quickly and accurately in this situation.

No sense of passage

In the second group, the testers were arranged to pass the terminal without stopping, and only turn their faces to the terminal; This test simulates the scene of multiple people continuously and quickly punch in and out, which requires higher response time. Moyan can still correctly recognize in a short time.

In the continuous recognition test, the average face recognition speed is about 30 ms, and the fastest recognition speed is less than 10 ms. The speed of the image equipment used in the evaluation is 10 pictures per second, and a single picture takes 1 / 10s. From the GIF image, as long as the face appears in the recognition range, Moyan can complete face recognition within 10ms-30ms, realizing the sensorless traffic.

Multi face recognition

When more than one person appears at the same time, the terminal will select the face with larger area in the picture for recognition, that is, when more than one person’s face appears at the same time, the terminal will select the nearest face for recognition, but it can’t detect and recognize all the faces appearing in the picture at the same time.

Security of live detection algorithm

In the security test, live photo deception test and live video deception test are used to test the security of the terminal.

In the photo test, the recent color certificate photo was selected to ensure the clarity of facial features; The video test uses the local short face video recorded in the scene environment to ensure the similarity of the environment.

Moyan’s support for live detection algorithm greatly improves the ability of face recognition to prevent malicious deception, and ensures the application security. In the security test, the photo deception all failed; In the video deception test, there is a very small probability of successful recognition; In principle, live recording face video for deception is set up to test the performance of the terminal. In reality, it needs to be recorded with high-definition devices, and it takes a lot of time to try. It is not a feasible means to break the defense, so there is no need to worry about this aspect of security when using it.

Evaluation summary

Face recognition has a strong demand on the scene. Whether the product can meet the needs of actual use, the core is not only the algorithm itself, but also the deep cultivation of the scene. The strong proof of the recognition rate of the algorithm level is only between the training set and the test set, which is the “theoretical value” existing in the laboratory.

In real life, there are a lot of uncontrollable factors in the process of face recognition. The direction and intensity of light, the change of beard and hairstyle, and even the expression will affect the recognition effect. Therefore, in order to accurately identify the scene, it is necessary to collect a large amount of scene data according to the different characteristics of the scene, continuously debug parameters, combine algorithms and methods, and even use peripheral hardware to improve the effect, and continuously iterate to achieve productization.

Moyan can adapt to more than 90% of indoor and outdoor environment applications, and in practical application, it has the characteristics of recognition rate higher than 99.9%, millisecond level fast recognition, support for live detection, anti backlight and so on. Local 1: n comparison, online witness comparison, real-time face detection and other functions make Moyan suitable for all kinds of places. At the same time, through the integration and encapsulation of algorithm, database and driver, and supporting the secondary development of SDK, it can become an open hardware platform for customers.

Summary of advantages

Rich interfaces, which can meet the needs of mainstream types of interfaces;

Domestic top industrial design, using aluminum alloy material, sandblasting oxidation manufacturing process;

It is easy to deploy and install, and the application function is perfect;

Support the secondary development of SDK;

Support 5000 to 10000 local face database recognition, the recognition speed is millisecond;

The accuracy of face recognition was 99.9%;

Reserve space at the bottom of the screen to display value-added advertising information;

It supports the function of person witness comparison and needs the cooperation of external card reader;

It supports the custom adjustment of the threshold and quality score of face recognition.

Space for improvement

Hardware: the teeth in the vertical joint of the panel are very small, so the angle can be adjusted accurately; The gear teeth in horizontal direction are large, so it is difficult to adjust the angle precisely. There are some inconveniences in the installation and angle adjustment.

Hardware: no physical switch is designed, and the working logic of the software is to directly enter the system or face recognition app by power on. In the case of a large number of deployment, it is impossible to start and stop the equipment in batch, so it is necessary to power on / off the terminals one by one.

Software: the logic of photo import and batch import is similar, and when using the batch import and photo import functions, the system automatically scans all the pictures in the external device for selection. It is not convenient to filter the pictures by importing them. It is recommended to add a file manager to manage the content of the external device.

Software: you can add the design of batch import and text pre entry Association, that is, after text pre entry, you can select and match from the entered text information when batch importing pictures, so as to complete personal information more quickly.

appendix

Product specifications

Principle of face recognition technology

Face recognition system mainly includes four parts: face image acquisition and detection, face image preprocessing, face image feature extraction and matching and recognition.

1. Face image collection: different face images can be collected through camera lens, such as static images, dynamic images, different positions, different expressions and so on. When the user is in the shooting range of the acquisition device, the acquisition device will automatically search and capture the user’s face image.

Face detection: in practice, face detection is mainly used for the preprocessing of face recognition, that is, to accurately calibrate the position and size of the face in the image. Face images contain a lot of pattern features, such as histogram features, color features, template features, structure features and Haar features. Face detection is to pick out the useful information and use these features to realize face detection.

2. Face image preprocessing: the face image preprocessing is based on the face detection results to process the image and finally serve the feature extraction process. Due to the limitation of various conditions and random interference, the original image obtained by the system can not be used directly, so it must be preprocessed in the early stage of image processing, such as gray correction, noise filtering and so on. For face image, the preprocessing process mainly includes light compensation, gray transformation, histogram equalization, normalization, geometric correction, filtering and sharpening.

3. Face image feature extraction: the features that can be used in face recognition system are usually divided into visual features, pixel statistical features, face image transform coefficient features, face image algebraic features, etc. Face feature extraction is based on some features of human face. Face feature extraction, also known as face representation, is the process of face feature modeling. Face feature extraction methods can be divided into two categories: one is knowledge-based representation method; The other is based on algebraic features or statistical learning.

The knowledge-based representation method is mainly based on the shape description of face organs and the distance between them to obtain feature data which is helpful for face classification. Its feature components usually include Euclidean distance, curvature and angle between feature points. Face is composed of eyes, nose, mouth, chin and other parts. The geometric description of these parts and the structural relationship between them can be used as important features for face recognition. These features are called geometric features. Knowledge based face representation mainly includes geometric feature-based method and template matching method.

Face image matching and recognition: the extracted feature data of face image is searched and matched with the feature template stored in the database. A threshold is set. When the similarity exceeds the threshold, the matching result is output. Face recognition is to compare the face features to be recognized with the face feature template, and judge the identity information of the face according to the similarity degree. This process can be divided into two categories: one is confirmation, which is the process of one-to-one image comparison; the other is identification, which is the process of one-to-many image matching and comparison.

Editor in charge: GT

Leave a Reply

Your email address will not be published. Required fields are marked *