HumanoID – The Results

Smart solution for verifying the identity of persons and validating the authenticity of official documents

Contact us
HumanoID - Estimated Results

Results Phase 1

In the first Phase multiple initial meetings were organized for commonly establish the next aspects of the project:

HumanoID - Estimated Results

Face Recognition

This module compares and validates the image of the face of a person from their CI document with a selfie image of the owner of the document

Liveness

The module that detects whether the face in a video is real (from a person present live at the capture point) or false (from a different source, in which the person is not present live).

OCR

The module that takes the image of the identity cards and extracts the information such as name, surname, domicile, personal number, series, number and validity period of the identity card, nationality, gender of the person, issuing section;

Results Phase 2

The running of the system presented in activity 2.1 is done on several instances, according to the type of the executed action. The web application has been configured on a cloud system, with 100% accessibility and uptime. Also, the data visualization application works in the same way. In order to limit the costs necessary for the maintenance of a system that can run the machine learning algorithms necessary for the application, the configuration of the algorithms, but also of the server that controls the access to the algorithms and the application flow was made on a local server.

The testing of the algorithms in activity 2.2 was done in a static environment, using recordings, so that it could be used in a dynamic environment, having a video stream, unlimited, with user interaction and providing guidance, completions for the extraction of partial results during the execution were brought and put in place, as well as their selection and validation during the completion of the flow. The system has the task of making a single informed decision on the authenticity of the person in question, so the 3 methods presented above will be used in a voting system, each model having a predetermined share, depending on the importance of the algorithm and its accuracy. analyzed in activity 2.2. Thus, based on this calculation, the confidence that the person is real will be established, and in the end, a decision will be made based on a lower credit limit.

Contact us

The main components that make up the purpose of the project, starting from the most abstract level to the user interface, are:

  • artificial intelligence algorithms and data analysis procedures
  • command server
  • web application
  • back office server and data visualization( back office)

According to the architecture presented in Activity 2.1, the algorithms used are divided into several categories defined by the function they perform in the system:

  • detection and formalization of the identity document
  • extracting information from the identity document
  • facial comparison between the identity document and a frame of the person in question detecting the authenticity of the person: false detection, head position detection, emotion detection

The testing of the experimental model and the prototype conformed to a software testing strategy on several levels of the application:

  • unit tests at the code level
  • integration tests to verify the performance of the evaluation of the results
  • evaluation tests – for choosing methods with greater accuracy on test data
  • performance-related performance tests

To ensure that the algorithms will have the same quality of performance, after being linked to the web application part, several tests were performed on the collected data set.

  • unit tests at the code level
  • integration tests to verify the performance of the evaluation of the results
  • evaluation tests – for choosing methods with greater accuracy on test data
  • performance-related performance tests

The whole enrollment process has five stages

presentation of the identity card at the video capture lens (webcam/phone camera)

video capture of the person looking at the camera lens

HumanoID - Estimated Results
HumanoID - Estimated Results

capturing a selfie picture

performing two head movements (two randomly selected left, right, up, and down movements)

the stage at which the user must smile broadly

The enrollment process explained

  • The first stage
  • The second stage
  • The third stage
  • The fourth stage
  • The fifth stage
  • End result

The quality of the information extracted by the OCR algorithm is closely related to the quality of the captured image. In case the quality is similar to the images captured by the mobile phone, the best results of the tested algorithms obtained the following accuracy scores on the collected data set: CNP 98%, name 94%, first name 92%, validity date 90 %, series 96%, number 94%, nationality 100%. For extremely low-quality images, in the case of very short bandwidth connections, the quality of the results decreases dramatically.

In the second stage, the user is informed that a selfie picture of his face will be captured, in order to be able to compare it with the face extracted from the identity card. The accuracy of the model is 99.5%.

The third stage is the false detection stage. At this stage, a video is taken with the user and forwarded for frame-by-frame analysis. In this case, an accuracy of 92% was recorded.

In stage four, the user will have to perform two randomly chosen head movements, left, right, up, and down. After two movements (out of the four listed above) are detected, or the time allowed per state is exceeded (one minute), a decision is made for this step: true if the algorithm predictions are the required random movements, in the given order, or false otherwise. The model is obtaining an accuracy of 90%.

In the fifth stage, the user must smile broadly. Each frame in the video stream is sent to the emotion detection algorithm, where a frame-by-frame prediction will be obtained. If most predictions are true (‘happiness’ emotion is detected), then the decision of the emotion detection algorithm will be true and false otherwise. The model offers an accuracy of 92%.

Once the user has gone through all five stages, depending on the prediction of the face recognition algorithm and the decision made by the voting system, enrollment can be successfully completed (if both are true) or not (if both are false).

Get in touch

We are always happy to talk

Phone

+40-21-223-7700

Email

hello@tremend.com

Address

165 Splaiul Unirii, Timpuri Noi Square,
TN Office 2 building, 4th floor,
District 3, Bucharest, Romania, 030134