Standards Development

Scope

The standard will establish requirements for automated systems that

• give scalar summarize the quality of face images that may be used in automated face recognition processes.

• measure specific properties of face images that are known to influence automated face recognition outcomes.

For summarization, the standard establishes requirements:

• for image quality assessment algorithms that compute one, or two, scalar quality value from an image.

• for image quality aggregation algorithms that summarize the quality of a large collection of images of different people

• on performance of image quality assessment algorithms

• on performance tests of image quality assessment algorithms For measurement of properties, the standard establishes

• The names of elements of an image quality vector (e.g. “occlusion”)

• The numerical definitions of elements in an image quality vector (e.g. “mis-focus”)

• Encodings of image quality vectors

• Requirements for aggregation of image quality vectors

• Requirements on performance tests of image quality assessment algorithms

The standard includes informative content on how quality algorithms may be used. It also includes recommendations on use of automated face recognition instead of quality assessment.

The standard is intended to support instantiation of fully conformant ISO/IEC 19794-5 and 39794-5 records that will be persist on credentials or in authoritative databases. It may additionally apply to collection of images in authentication or identification attempts also.

The standard is expected to include definitions of the Level 3 image appearance conformance tests declared but not implemented in ISO/IEC 29109-5 i.e. tests of requirements written in ISO/IEC 19794-5 (and ISO/IEC 39794-5) that require image analysis to perform (e.g. head yaw estimation).

The standard does not establish requirements on cameras and imaging systems.

The standard could be progressed as part of a revision ISO/IEC TR 29794-5:2010 Information technology -- Biometric sample quality -- Part 5: Face image data

Purpose

Face recognition tests reveal that error rates are non-zero. Particularly face recognition algorithms sometimes yield low similarity scores when comparing an image with a prior image. One cause of this is poor image quality. The term quality here is used to refer to a general lack of information in an image, e.g. due to blur, or lack of suitability for recognition e.g. non-frontal view. Another cause is the change of appearance as the face ages but that is a separate issue and not the subject of this standard.

This standard is intended to support high accuracy face recognition by providing numerical quality values

• against which a photograph may be accepted or rejected, or selected from a set of images. Its primary role is at the time of capture so that re-capture can be attempted.

• against which a collection sub-system can provide actionable feedback to a subject, photographer or attending official who might then initiate collection of a replacement photo.

High performance implementations of this standard should:

• Accurately measure relevant image properties.

• Allow system owners to identify particular locations, time periods, populations, or circumstances for which specific image quality defects are common or trending e.g. in under-exposure of the face. High performance implementations of this standard should:

• Reject images whose properties present a risk of false rejection from a generic automated recognition algorithms;

• Reject images for which the face in the image is presented in a manner that may cause false rejection from automated recognition algorithms.

• Allow system owners to identify particular locations, time periods, populations, or circumstances for which image quality is poor in aggregate. This can be achieved by aggregating scores over collections of images.

An image quality assessment algorithm may be used alone, or in conjunction with the review of trained human adjudicator, or in conjunction with an algorithm that makes numerical measurements of specific face image properties such as head orientation, illumination uniformity, and expression neutrality.

Image quality assessment algorithms may additionally support accurate face analysis e.g. age estimation.