GB/T 33767.5-2018Information technology -- Biometric sample quality -- Part 5: Face image data (English PDF)

7.1 Overview

Different factors should be considered when analyzing face image quality. These factors can be divided into.

a) Image attributes, such as the size or resolution of the image;

b) Image appearance characteristics, such as exposure or noise;

c) Environmental characteristics, such as lighting or background;

d) Features such as the consistency between the skin color displayed in the image and the subject's skin color;

e) The subject's behavior.

Some of the above attributes and characteristics are difficult to evaluate and evaluate, such as the matching degree between the skin color displayed in the image and the subject's skin color.

See ISO /IEC 19794-5 for requirements of some attributes and characteristics such as eye distance (in pixels). The evaluation of these attributes and characteristics requires

More complex algorithms and technologies in computer vision and image understanding. In addition, there may be different processing methods, such as automatic detection

The position of the eyes in the face image is based on various principles. According to GB/T 33767.1, the regularized quality score can be obtained.

The FQAA algorithm can detect images without segmenting the face area (for example, when measuring the image size to evaluate the static of the acquisition process).

State features, such as compression rate, compression method, sensor resolution) or only analyze the face area (for example, estimate the pose of the subject). Locally

In the region, the local structure of the face is defined by the pixel value (original or processed); a single quality score may be merged from multiple local results.

Various face image quality evaluation algorithms can be developed for different environments, cameras, subjects and other factors, which are shown on different data sets.

Different performance.

7.2 Dynamic subject characteristics

7.2.1 Subject behavior

Common characteristics related to the subject's behavior include.

---Open your eyes;

---Open mouth;

---Various expressions, such as smiling or neutral;

---Head posture, such as facing directly or turning to any direction.

Similar to environmental attributes or features, the quantification of the above parameters requires recognition of background, face and face features.

In addition, the core algorithms (evaluating certain attributes) in the algorithm are required to be implemented on a computer, and the core algorithms can be obtained.

The quantitative value of the evaluation performance, in order to achieve the purpose of reducing the computational complexity. If conditions permit, you can select the most commonly used algorithm or

Read to reduce complexity.

7.2.2 Analysis based on the statistical difference between the left half of the face and the right half of the face

7.2.2.1 Illumination symmetry

Assuming a two-dimensional portrait image (see the image specified in ISO /IEC 19794-5.2005/Amd.1), it can be divided by left and right symmetry.

Analyze the quality of lighting and posture. The face area is divided into left and right areas based on the center line of the eyes (Figure 1). The following symmetry analysis is

Detect the difference between the corresponding areas on the left and right of the face. The difference value represents the symmetry of a certain local image attribute, such as the original pixel value or the local filtered image.

Prime value. The local image filter can use Gabor filter, local binary filter, sequential filter or any other suitable

Local filter. The difference between the left and right regions gives the light quality score (i.e. how symmetrical the light is) or the attitude quality score (i.e. the frontal attitude).

how). Most faces are symmetrical, but some people have obvious differences between the left and right parts of the face, such as marks, discoloration, etc.

different. The symmetry of quality analysis measures should consider these differences.

Note. Figure 1 is from the face image database, see reference [18].

The difference in illumination symmetry can be based on some local feature histograms HLm*n and HRm*n of the left and right face regions, where m is the feature vector

The dimension of, n is the number of histogram groups. The histogram difference calculation formula is as follows.

Di= HLm*n-HRm*n (1)

In formula (1) |.| is a suitable form of histogram distance, such as histogram intersection, cross entropy or KL distance. The greater the difference, the greater the

The worse the left-right symmetry, the lower the image quality in some respects. One possibility is to use the image to regularize the pixel values.

A.1 gives an analysis example.

7.2.2.2 Attitude symmetry

Pose symmetry is an important factor in the judgment of face image quality. The direction of face rotation is divided into up and down rotation angle (pitch) and depth rotation angle

(yaw) and in-plane rotation angle (rol) three components, of which the change of the depth rotation angle component is the most important factor affecting the symmetry of the attitude

Vegetarian. It is generally considered that a face image with a depth rotation angle component between -5° and 5° is a high-quality image, and the depth rotation angle component is between -30° and

A face image between 30° is an acceptable image.

The posture symmetry analysis should be based on the posture-sensitive image attributes. The local binary mode (LBP) can be used to filter pixel values.

A.2 gives an analysis example.

7.3 Static characteristics of the acquisition process

7.3.1 Overview

Typical scene characteristics describing environmental impact are as follows.

---Image enhancement and data compression process, such as image resolution and size;

---Static camera features, such as. resolution;

---Static features of the background, such as wallpaper.

According to attributes or characteristics, quantification of these parameters requires the identification of background, face and facial features.

Different core algorithms and their performance values can be used here.

7.3.2 Image resolution and size

The number of image rows and columns of pixels can be used to indicate the nominal resolution. The interpupillary distance in pixels can be used to measure relative to the face

The pixel range of the feature. In addition, using the statistical average of the interpupillary distance (for example, 63mm), the pixel density can be converted into a spatial sampling rate.

7.3.3 Noise

7.3.3.1 Noise source

The noise in the face image comes from the various processes of acquiring the digital image. Different equipment or processes correspond to different introduced noises. phase

Noise sources include.

---Digital image acquisition equipment, such as the image sensor of a digital camera;

---Analog image acquisition equipment;

---Image scanning equipment;

---Image compression algorithm, such as JPEG or wavelet compression.

7.3.3.2 Image acquisition noise

The upper limit of image noise can be estimated by using the segmented smooth image prior model and measuring the response function of the CCD camera.

7.3.3.3 DCT compression noise

Compressed artifacts caused by DCT compression can be estimated by the difference between the pixel gradient across the block boundary and the gradient of the internal pixels.

To calculate the discontinuity of the boundary, the method of measuring the noise in the block can also be used.

7.4 Image acquisition characteristics

7.4.1 Image properties

Metadata can be stored in the image according to different standards.

Note. For example, digital cameras can store metadata information in EXIF format. EXIF is a "standard of image file format used by digital cameras".

EXIF information contains the state of the camera when the information was captured. And some of the information (such as exposure time) is definitely helpful for quality assessment. in

In the EXIF standard, different metadata tags are defined such as. data and time information, camera settings, exposure, location information, description and copyright information.

The verification of image attributes does not need to rely on the metadata contained in the image source.

7.4.2 Image appearance

The appearance of the image should depend on the color distribution of the image.