QualityAssessment

We are working on the implementation of a Face Image Quality Assessment according to ISO/IEC 29794-5.

Message definition

message QualityAssessment {
    string check = 1;
    double score = 2;
    string message = 3;
}

QualityAssessment fields

check

The quality check performed, see below.

score

The outcome of the quality check. This is a score int the range between 0.0 and 1.0. The higher the value, the better the check was passed. Some tests simply return 0.0 or 1.0 here, if the check fails or succeeds respectively.

message

A text with additional info about this quality assessment.

Quality Checks performed

ImageSize

The size of any input image should not be bigger than the recommended image size (1200 x 1600). If an image is bigger than this size in both dimensions, it is automatically scaled down.
This check is only reported in case that the image had to be scaled down and is just for your information.

SuitableFace

The service typically needs to find at least one suitable face in the image. Not all found faces are suitable to be used by a specific API. In case that a found face is not suitable it will be rejected and a quality assessment message is generated containing the reason for the rejection.
This check is only reported in case that a non suitable face was found and is just for your information.
If all found faces have been rejected, an additional FaceNotFound error will be generated of course.

Possible reasons for non suitable faces are:

• EyesInverted: the eyes are inverted, i.e. the face seems to be upside down.
• NoseAboveEyes: the nose is above the eyes.
• (Profile: seems to be a profile face image.)
• EyesTooCloseTogether: minimum horizontal eye distance not reached.
• BeyondBorder: at least one of the eyes is too close to the border.

FullyVisibleFace

To crop out the face-area of a found face from an image, there needs to be enough space around the face. Especially if the face is tilted and needs to be rotated, more space around the face is needed.
This check is only reported in case that a face is too close to the edge and cannot be completely cropped.
If no face is fully visible within an image, an additional ThumbnailExtractionFailed error will be generated.

OFIQ Quality Measures

In addition to our custom checks, the BWS 3 implements the Open Source Face Image Quality (OFIQ) library. OFIQ is the reference implementation for the ISO/IEC 29794-5 standard. For more information, see the official OFIQ GitHub Project.

The table below lists the 28 OFIQ quality measures performed by our service: