Facial Expression Recognition (Face Recognition Techniques) Part 1

Introduction

Facial expressions are the facial changes in response to a person’s internal emotional states, intentions, or social communications. Facial expression analysis has been an active research topic for behavioral scientists since the work of Darwin in 1872 [21, 26, 29, 83]. Suwa et al. [90] presented an early attempt to automatically analyze facial expressions by tracking the motion of 20 identified spots on an image sequence in 1978. After that, much progress has been made to build computer systems to help us understand and use this natural form of human communication [5, 7, 8, 17, 23, 32, 43, 45, 57, 64, 77, 92, 95, 106-108, 110].

In this topic, facial expression analysis refers to computer systems that attempt to automatically analyze and recognize facial motions and facial feature changes from visual information. Sometimes the facial expression analysis has been confused with emotion analysis in the computer vision domain. For emotion analysis, higher level knowledge is required. For example, although facial expressions can convey emotion, they can also express intention, cognitive processes, physical effort, or other intra- or interpersonal meanings. Interpretation is aided by context, body gesture, voice, individual differences, and cultural factors as well as by facial configuration and timing [11, 79, 80]. Computer facial expression analysis systems need to analyze the facial actions regardless of context, culture, gender, and so on.


Basic structure of facial expression analysis systems

Fig. 19.1 Basic structure of facial expression analysis systems

The accomplishments in the related areas such as psychological studies, human movement analysis, face detection, face tracking, and recognition make the automatic facial expression analysis possible. Automatic facial expression analysis can be applied in many areas such as emotion and paralinguistic communication, clinical psychology, psychiatry, neurology, pain assessment, lie detection, intelligent environments, and multimodal human computer interface (HCI).

Principles of Facial Expression Analysis

Basic Structure of Facial Expression Analysis Systems

Facial expression analysis includes both measurement of facial motion and recognition of expression. The general approach to automatic facial expression analysis (AFEA) consists of three steps (Fig. 19.1): face acquisition, facial data extraction and representation, and facial expression recognition.

Face acquisition is a processing stage to automatically find the face region for the input images or sequences. It can be a detector to detect face for each frame or just detect face in the first frame and then track the face in the remainder of the video sequence. To handle large head motion, the head finder, head tracking, and pose estimation can be applied to a facial expression analysis system.

After the face is located, the next step is to extract and represent the facial changes caused by facial expressions. In facial feature extraction for expression analysis, there are mainly two types of approaches: geometric feature-based methods and appearance-based methods. The geometric facial features present the shape and locations of facial components (including mouth, eyes, brows, nose, etc.). The facial components or facial feature points are extracted to form a feature vector that represents the face geometry. With appearance-based methods, image filters, such as Gabor wavelets, are applied to either the whole-face or specific regions in a face image to extract a feature vector. Depending on the different facial feature extraction methods, the effects of in-plane head rotation and different scales of the faces can be eliminated by face normalization before the feature extraction or by feature representation before the step of expression recognition.

Facial expression recognition is the last stage of AFEA systems. The facial changes can be identified as facial action units or prototypic emotional expressions (see Sect. 19.3.1 for definitions). Depending on whether the temporal information is used, in this topic we classify a recognition approach as frame-based or sequence-based.

Emotion-specified facial expression (posed images from database [49]). 1, disgust; 2, fear; 3, joy; 4, surprise; 5, sadness; 6, anger

Fig. 19.2 Emotion-specified facial expression (posed images from database [49]). 1, disgust; 2, fear; 3, joy; 4, surprise; 5, sadness; 6, anger

Organization of the topic

This topic introduces recent advances in facial expression analysis. The first part discusses general structure of AFEA systems. The second part describes the problem space for facial expression analysis. This space includes multiple dimensions: level of description, individual differences in subjects, transitions among expressions, intensity of facial expression, deliberate versus spontaneous expression, head orientation and scene complexity, image acquisition and resolution, reliability of ground truth, databases, and the relation to other facial behaviors or nonfacial behaviors. We note that most work to date has been confined to a relatively restricted region of this space. The last part of this topic is devoted to a description of more specific approaches and the techniques used in recent advances. They include the techniques for face acquisition, facial data extraction and representation, facial expression recognition, and multimodal expression analysis. The topic concludes with a discussion assessing the current status, future possibilities, and open questions about automatic facial expression analysis.

Problem Space for Facial Expression Analysis

Level of Description

With few exceptions [17, 23, 34, 95], most AFEA systems attempt to recognize a small set of prototypic emotional expressions as shown in Fig. 19.2 (i.e., disgust, fear, joy, surprise, sadness, anger). This practice may follow from the work of Darwin [21] and more recently Ekman and Friesen [27, 28] and Izard et al. [48] who proposed that emotion-specified expressions have corresponding prototypic facial expressions. In everyday life, however, such prototypic expressions occur relatively infrequently. Instead, emotion more often is communicated by subtle changes in one or a few discrete facial features, such as tightening of the lips in anger or obliquely lowering the lip corners in sadness [12]. Change in isolated features, especially in the area of the eyebrows or eyelids, is typical of paralinguistic displays; for instance, raising the brows signals greeting [25]. To capture such subtlety of human emotion and paralinguistic communication, automated recognition of fine-grained changes in facial expression is needed. The facial action coding system (FACS: [29]) is a human-observer-based system designed to detect subtle changes in facial features. Viewing videotaped facial behavior in slow motion, trained observers can manually FACS code all possible facial displays, which are referred to as action units and may occur individually or in combinations.

Table 19.1 FACS action units (AU). AUs with “*” indicate that the criteria have changed for this AU, that is, AU 25, 26, and 27 are now coded according to criteria of intensity (25A-E), and AU 41, 42, and 43 are now coded according to criteria of intensity

FACS action units (AU). AUs with “*” indicate that the criteria have changed for this AU, that is, AU 25, 26, and 27 are now coded according to criteria of intensity (25A-E), and AU 41, 42, and 43 are now coded according to criteria of intensity

FACS consists of 44 action units. Thirty are anatomically related to contraction of a specific set of facial muscles (Table 19.1) [22]. The anatomic basis of the remaining 14 is unspecified (Table 19.2). These 14 are referred to in FACS as miscellaneous actions. Many action units may be coded as symmetrical or asymmetrical. For action units that vary in intensity, a 5-point ordinal scale is used to measure the degree of muscle contraction. Table 19.3 shows some examples of combinations of FACS action units.

AU

Description

8

Lips toward

19

Tongue show

21

Neck tighten

29

Jaw thrust

30

Jaw sideways

31

Jaw clench

32

Bite lip

33

Blow

34

Puff

35

Cheek suck

36

Tongue bulge

37

Lip wipe

38

Nostril dilate

39

Nostril compress

Table 19.2 Miscellaneous actions

Table 19.3 Some examples of combination of FACS action units

Some examples of combination of FACS action units

Although Ekman and Friesen proposed that specific combinations of FACS action units represent prototypic expressions of emotion, emotion-specified expressions are not part of FACS; they are coded in separate systems, such as the emotional facial action system (EMFACS) [41]. FACS itself is purely descriptive and includes no inferential labels. By converting FACS codes to EMFACS or similar systems, face images may be coded for emotion-specified expressions (e.g., joy or anger) as well as for more molar categories of positive or negative emotion [65].

Individual Differences in Subjects

Face shape, texture, color, and facial and scalp hair vary with sex, ethnic background, and age [33, 119]. Infants, for instance, have smoother, less textured skin and often lack facial hair in the brows or scalp. The eye opening and contrast between iris and sclera differ markedly between Asians and Northern Europeans, which may affect the robustness of eye tracking and facial feature analysis more generally. Beards, eyeglasses, or jewelry may obscure facial features. Such individual differences in appearance may have important consequences for face analysis. Few attempts to study their influence exist. An exception was a study by Zlochower et al. [119], who found that algorithms for optical flow and high-gradient component detection that had been optimized for young adults performed less well when used in infants. The reduced texture of infants’ skin, their increased fatty tissue, juvenile facial conformation, and lack of transient furrows may all have contributed to the differences observed in face analysis between infants and adults.

In addition to individual differences in appearance, there are individual differences in expressiveness, which refers to the degree of facial plasticity, morphology, frequency of intense expression, and overall rate of expression. Individual differences in these characteristics are well established and are an important aspect of individual identity [61] (these individual differences in expressiveness and in biases for particular facial actions are sufficiently strong that they may be used as a biometric to augment the accuracy of face recognition algorithms [19]). An extreme example of variability in expressiveness occurs in individuals who have incurred damage either to the facial nerve or central nervous system [75, 99]. To develop algorithms that are robust to individual differences in facial features and behavior, it is essential to include a large sample of varying ethnic background, age, and sex, which includes people who have facial hair and wear jewelry or eyeglasses and both normal and clinically impaired individuals.

Transitions Among Expressions

A simplifying assumption in facial expression analysis is that expressions are singular and begin and end with a neutral position. In reality, facial expression is more complex, especially at the level of action units. Action units may occur in combinations or show serial dependence. Transitions from action units or combination of actions to another may involve no intervening neutral state. Parsing the stream of behavior is an essential requirement of a robust facial analysis system, and training data are needed that include dynamic combinations of action units, which may be either additive or nonadditive.

As shown in Table 19.3, an example of an additive combination is smiling (AU 12) with mouth open, which would be coded as AU 12 + 25, AU 12 + 26, or AU 12 + 27 depending on the degree of lip parting and whether and how far the mandible was lowered. In the case of AU 12 + 27, for instance, the facial analysis system would need to detect transitions among all three levels of mouth opening while continuing to recognize AU 12, which may be simultaneously changing in intensity.

Nonadditive combinations represent further complexity. Following usage in speech science, we refer to these interactions as co-articulation effects. An example is the combination AU 12 +15, which often occurs during embarrassment. Although AU 12 raises the cheeks and lip corners, its action on the lip corners is modified by the downward action of AU 15. The resulting appearance change is highly dependent on timing. The downward action of the lip corners may occur simultaneously or sequentially. The latter appears to be more common [85]. To be comprehensive, a database should include individual action units and both additive and nonadditive combinations, especially those that involve co-articulation effects. A classifier trained only on single action units may perform poorly for combinations in which co-articulation effects occur.

Intensity of Facial Expression

Facial actions can vary in intensity. Manual FACS coding, for instance, uses a 3- or more recently a 5-point intensity scale to describe intensity variation of action units (for psychometric data, see Sayette et al. [82]). Some related action units, moreover, function as sets to represent intensity variation. In the eye region, action units 41, 42, and 43 or 45 can represent intensity variation from slightly drooped to closed eyes. Several computer vision researchers proposed methods to represent intensity variation automatically. Essa and Pentland [32] represented intensity variation in smiling using optical flow. Kimura and Yachida [50] and Lien et al. [56] quantified intensity variation in emotion-specified expression and in action units, respectively. These authors did not, however, attempt the more challenging step of discriminating intensity variation within types of facial actions. Instead, they used intensity measures for the more limited purpose of discriminating between different types of facial actions. Tian et al. [94] compared manual and automatic coding of intensity variation. Using Gabor features and an artificial neural network, they discriminated intensity variation in eye closure as reliably as human coders did. Recently, Bartlett and colleagues [5] investigated action unit intensity by analyzing facial expression dynamics. They performed a correlation analysis to explicitly measure the relationship between the output margin of the learned classifiers and expression intensity. Yang et al. [111] converted the problem of intensity estimation to a ranking problem, which is modeled by the RankBoost. They employed the output ranking score for intensity estimation. These findings suggest that it is feasible to automatically recognize intensity variation within types of facial actions. Regardless of whether investigators attempt to discriminate intensity variation within facial actions, it is important that the range of variation be described adequately. Methods that work for intense expressions may generalize poorly to ones of low intensity.

Deliberate Versus Spontaneous Expression

Most face expression data have been collected by asking subjects to perform a series of expressions. These directed facial action tasks may differ in appearance and timing from spontaneously occurring behavior [30]. Deliberate and spontaneous facial behavior are mediated by separate motor pathways, the pyramidal and extrapyramidal motor tracks, respectively [75]. As a consequence, fine-motor control of deliberate facial actions is often inferior and less symmetrical than what occurs spontaneously. Many people, for instance, are able to raise their outer brows spontaneously while leaving their inner brows at rest; few can perform this action voluntarily. Spontaneous depression of the lip corners (AU 15) and raising and narrowing the inner corners of the brow (AU 1 + 4) are common signs of sadness. Without training, few people can perform these actions deliberately, which incidentally is an aid to lie detection [30]. Differences in the temporal organization of spontaneous and deliberate facial actions are particularly important in that many pattern recognition approaches, such as hidden Markov modeling, are highly dependent on the timing of the appearance change. Unless a database includes both deliberate and spontaneous facial actions, it will likely prove inadequate for developing face expression methods that are robust to these differences.

Head Orientation and Scene Complexity

Face orientation relative to the camera, the presence and actions of other people, and background conditions may influence face analysis. In the face recognition literature, face orientation has received deliberate attention. The FERET database [76], for instance, includes both frontal and oblique views, and several specialized databases have been collected to try to develop methods of face recognition that are invariant to moderate change in face orientation [100]. In the face expression literature, use of multiple perspectives is rare; and relatively less attention has been focused on the problem of pose invariance. Most researchers assume that face orientation is limited to in-plane variation [3] or that out-of-plane rotation is small [57, 68, 77, 95]. In reality, large out-of-plane rotation in head position is common and often accompanies change in expression. Kraut and Johnson [54] found that smiling typically occurs while turning toward another person. Camras et al. [10] showed that infant surprise expressions often occur as the infant pitches her head back. To develop pose invariant methods of face expression analysis, image data are needed in which facial expression changes in combination with significant non-planar change in pose. Some efforts have been made to handle large out-of-plane rotation in head position [5, 20, 97, 104].

Scene complexity, such as background and the presence of other people, potentially influences accuracy of face detection, feature tracking, and expression recognition. Most databases use image data in which the background is neutral or has a consistent pattern and only a single person is present in the scene. In natural environments, multiple people interacting with each other are likely, and their effects need to be understood. Unless this variation is represented in training data, it will be difficult to develop and test algorithms that are robust to such variation.

Image Acquisition and Resolution

The image acquisition procedure includes several issues, such as the properties and number of video cameras and digitizer, the size of the face image relative to total image dimensions, and the ambient lighting. All of these factors may influence facial expression analysis. Images acquired in low light or at coarse resolution can provide less information about facial features. Similarly, when the face image size is small relative to the total image size, less information is available. NTSC cameras record images at 30 frames per second, The implications of down-sampling from this rate are unknown. Many algorithms for optical flow assume that pixel displacement between adjacent frames is small. Unless they are tested at a range of sampling rates, the robustness to sampling rate and resolution cannot be assessed.

Within an image sequence, changes in head position relative to the light source and variation in ambient lighting have potentially significant effects on face expression analysis. A light source above the subject’s head causes shadows to fall below the brows, which can obscure the eyes, especially for subjects with pronounced bone structure or hair. Methods that work well in studio lighting may perform poorly in more natural lighting (e.g., through an exterior window) when the angle of lighting changes across an image sequence. Most investigators use single-camera setups, which is problematic when a frontal orientation is not required. With image data from a single camera, out-of-plane rotation may be difficult to standardize. For large out-of-plane rotation, multiple cameras may be required. Multiple camera setups can support three dimensional (3D) modeling and in some cases ground truth with which to assess the accuracy of image alignment. Pantic and Rothkrantz [70] were the first to use two cameras mounted on a headphone-like device; one camera is placed in front of the face and the other on the right side of the face. The cameras are moving together with the head to eliminate the scale and orientation variance of the acquired face images.

Image resolution is another concern. Professional grade PAL cameras, for instance, provide very high resolution images. By contrast, security cameras provide images that are seriously degraded. Although postprocessing may improve image resolution, the degree of potential improvement is likely limited. Also the effects of post processing for expression recognition are not known. Table 19.4 shows a face at different resolutions. Most automated face processing tasks should be possible for a 69 x 93 pixel image. At 48 x 64 pixels the facial features such as the corners of the eyes and the mouth become hard to detect. The facial expressions may be recognized at 48 x 64 and are not recognized at 24 x 32 pixels. Algorithms that work well at optimal resolutions of full face frontal images and studio lighting can be expected to perform poorly when recording conditions are degraded or images are compressed. Without knowing the boundary conditions of face expression algorithms, comparative performance is difficult to assess. Algorithms that appear superior within one set of boundary conditions may perform more poorly across the range of potential applications. Appropriate data with which these factors can be tested are needed.

 A face at different resolutions. All images are enlarged to the same size. At 48 x 64 pixels the facial features such as the corners of the eyes and the mouth become hard to detect. Facial expressions are not recognized at 24 x 32 pixels [97]

Table 19.4 A face at different resolutions. All images are enlarged to the same size. At 48 x 64 pixels the facial features such as the corners of the eyes and the mouth become hard to detect. Facial expressions are not recognized at 24 x 32 pixels [97]

Next post:

Previous post: