Face Verification

Face verification is the task of comparing a face image with another and verifying whether the two images match. The mapping is usually one-to-one. Face verification is different from face detection. In face detection, the model identifies the presence of a human or human-like face in an image without identifying the person.
In Rubiscape,, before face verification, face detection is carried out. Using the pre-built models, the face detection step detects all the faces in an image. Then, in the Face Verification step, the Source Image (also called the Input Image) and the Target Image are compared to determine whether they are the same or different.

Notes:

The Source and Target Images should not be the same. It means that the same file should not be used as the source and target image since verifying a person's identity from two same images would make no sense.

Models and Parameters used in Face Verification

There are several pre-trained models and other parameters required for face verification. Also, face verification is preceded by face detection. Listed below are the models and parameters used in Rubiscape for face verification.
For Face Detection,

MTCNN (Multi-task Cascaded Convolutional Neural Networks)
Retina Face (default)

For Face Verification,

Arc Face (default)
VGG-Face
Facenet
Facenet512
OpenFace
DeepFace

Also, we select a distance metric learning method from the following list of Distance Metrics.

Cosine Similarity (default)
Euclidean Distance
Euclidean_l2

Notes:

MTCNN is a faster algorithm compared to RetinaFace. It can be used in the case of images with many faces and where high accuracy of face detection is not expected.
MTCNN shows a lower accuracy in the case of tilted or distorted faces in images.
RetinaFace algorithm takes time to load but is more accurate with the face detection results.
From Face Detection results, you can download files for detected images and faces to your machine using the links provided in the output.
Arc Face is one of the newest face verification models and demonstrates 99.40% accuracy. It takes the two images as input and gives the distance between the two images as the output. This distance is then used to infer if the two faces belong to the same person.
VGG-Face (Visual Geometry Group - Face) is based on deep convolutional neural networks. It is known to demonstrate an accuracy of more than 97%.
Facenet and Facenet512 (by Google) are face recognition models that extract face embeddings. Face embeddings are high-quality face features to train and build a face identification system. The Facenet models demonstrate over 99.6% accuracy for face verification. Facenet is used for face verification, face recognition, and face clustering (grouping photos of people with the same identity).
OpenFace model is lightweight and has more license-type flexibility than the Facenet model. It achieves an accuracy close to 94% during face verification.
DeepFace (by Facebook) is a face recognition approach based on a deep neural network with nine layers. It achieves more than 97% accuracy and is very close to humans in recognizing faces.
The distance value indicates whether the model is successful in verifying the face. The lesser the distance value, the closer are the data points, and the face is similar. Conversely, the larger the distance, the farther are the data points, and the faces are dissimilar. |

The available properties of Face Verification are as shown in the figure below.

Using the Properties pane, you can select

A suitable name for the task
The Source Image
The Target Image
The Model for Detection is the pre-trained model for Face Detection
The Model for Verification is the pre-trained model for Face Verification
The Distance Metric is the metric to calculate the similarity between data points.

In addition, we write a code in RubiPython to

Create Custom Output Variables corresponding to the images used for analysis
Call the images from the database for analysis

A sample code in RubiPython is shown below.

A typical algorithm flow for Face Verification looks like the one shown below.

On exploring the Face Verification node, you see the results of the face verification analysis of the source and target images by the selected model.
For the images above,

The highlighted faces indicate that they are successfully detected.
Maximum Threshold To Verify = 0.68
Model: ArcFace
Similarity Metric = cosine

The image below shows the results of face verification analysis for the given pair of images, that is, inputImage1_Face1 and targetImage2_Face1.

The results indicate that.

Distance = 0.5115 and
Verified = true
The distance is less than the threshold, which means that the model successfully verifies that the two faces belong to the same person.

The image below shows the results of face verification analysis for a different pair of images, that is, inputImage1_Face1 and targetImage1_Face1.

The results indicate that,

Distance = 0.8768 and
Verified = false
The distance is more than the threshold, which means that the model successfully verifies that the two faces do not belong to the same person.

An interesting case of face verification is given below. Here, the face detection model detects a face-like structure of a wall hanging (in reality). Then the face verification model compares it with the target image.

The results indicate that,

Distance = 0.9153 and
Verified = false
The distance is more than the threshold (maximum among the three cases) which means that the model successfully verifies that the two faces do not belong to the same person.