Face Recognition Features

A broad definition of face recognition encompasses a variety of technologies used to extract data about facial images. Sony's face recognition technology offers the following features.

Face detection:

The detection and indication of facial zones that are facing in various directions in complex scenes.

Facial pose estimation:

The estimation of the direction (angle) to which a face is turned

Facial part detection:

The identification of the positions of facial parts such as the centers of the eyes, the tip of the nose, and the corners of the mouth

Facial attribute classification:

The classification of faces by gender, ethnicity, age, expression and other characteristics

Face identification:

The identification of individuals through comparisons with registered people (This is the narrow definition of face recognition.)

Multi-view face detection:

In this example, even faces that are not turned toward the front have been detected, and facial poses have been estimated at the same time. The direction of the arrows indicates the direction of the face, and the length of the arrows indicates the value of angle. (An angle of zero denotes a frontal pose, while a larger angle indicates that the face is turned to one side.)

Multi-view face detection: In this example, even faces that are not turned toward the front have been detected, and facial poses have been estimated at the same time. The direction of the arrows indicates the direction of the face, and the length of the arrows indicates the value of angle. (An angle of zero denotes a frontal pose, while a larger angle indicates that the face is turned to one side.)

Detecting facial parts

Detecting facial attributes: This example shows how the system can discriminate between males and females, adults and children, and smiling and non-smiling faces. The length of the bars indicates the classification scores.

Face identification: In this example, the system has identified two registered persons and displayed their names, while determining that the third person is not registered.

Statistical Face Recognition

The following description relates to the type of face recognition that is most commonly used in commercial applications.

The first step is to define a facial pattern of a specific size. Human vision can judge whether or not a face is present even in a low-resolution image made up of 16x16 pixels. This ability does not rely on color, and human eyes will find faces even in a monochrome image. Computers process facial patterns using images of about the same size.

1. Detection of face to be scanned

The system scans the image from top left to bottom right until it finds this pattern.

2. Facial pattern classification

A PC cluster system

Facial patterns are not easy to define. They vary from person to person, and they also change according to the angle of the face and differences in lighting conditions or facial expressions. To overcome this, it is necessary to formulate functions that allow discrimination between facial and non-facial images by applying statistical methods to large numbers of facial and non-facial images. The key to this is the use of features. To facilitate the implementation of this technology on consumer electronic products, Sony uses a unique set of extremely simple features. Because these features can be combined in various ways, it is possible to achieve powerful pattern classification performance despite the simplicity of the operations involved.

Analysis takes longer. However, by using a PC cluster system installed for use in artificial intelligence research, Sony has been able to detect optimal features and compile a compact dictionary of facial patterns.