Robust representations for face recognition: the power of averages

We are able to recognise familiar faces easily across large variations in image quality, though our ability to match unfamiliar faces is strikingly poor. Here we ask how the representation of a face changes as we become familiar with it. We use a simple image-averaging technique to derive abstract representations of known faces. Using Principal Components Analysis, we show that computational systems based on these averages consistently outperform systems based on collections of instances. Furthermore, the quality of the average improves as more images are used to derive it. These simulations are carried out with famous faces, over which we had no control of superficial image characteristics. We then present data from three experiments demonstrating that image averaging can also improve recognition by human observers. Finally, we describe how PCA on image averages appears to preserve identity-specific face information, while eliminating non-diagnostic pictorial information. We therefore suggest that this is a good candidate for a robust face representation.     

熟悉面孔与陌生面孔识别的差异

摘 要：面孔识别包括熟悉面孔识别和陌生面孔识别,两者之间的差异已经得到了多方面的证明。神经生理学的研究发现,识别陌生面孔和熟悉面孔所激活的脑区存在差异,与识别陌生面孔相比,识别熟悉面孔的脑区活动范围更大,包括双侧前额叶、单侧颞叶、海马回、杏仁核、后扣带回和右下额叶。并且也存在脑电的差异,比如 N400和P600的差异。对陌生面孔和熟悉面孔识别具有相同影响的因素有：遮挡、倒置等;具有不同影响的因素有：视角、表情、内外部特征等。文章也探讨了陌生面孔变成熟悉面孔的过程,以及在这一过程中起主要作用的编码形式等。     

Hemispheric asymmetry in cross-race face recognition

This study examines two phenomena related to face perception, both of which depend on experience and holistic processing: perceivers process faces more efficiently in the right hemisphere of the brain (a hemispheric asymmetry), and they typically show greater recognition accuracy for members of their racial ingroup (a cross-race recognition deficit). The current study tests the possibility that these two effects are related. If asymmetry depends on experience, it should be particularly evident with (more familiar) ingroup faces; if cross-race recognition relies on holistic processing, it should be particularly evident for faces presented to the right hemisphere. Black and White participants viewed Black and White faces presented to either the left or right visual field. As predicted, participants showed a more pronounced asymmetry for ingroup (rather than outgroup) faces, and cross-race recognition deficits were more pronounced for stimuli presented to the left (rather than the right) visual field.     

Famous face recognition,face matching,and extraversion

It has been previously established that extraverts who are skilled at interpersonal interaction perform significantly better than introverts on a face-specific recognition memory task. In our experiment we further investigate the relationship between extraversion and face recognition, focusing on famous face recognition and face matching. Results indicate that more extraverted individuals perform significantly better on an upright famous face recognition task and show significantly larger face inversion effects. However, our results did not find an effect of extraversion on face matching or inverted famous face recognition.     

Tolerance of geometric distortions in infant's face recognition

The aim of the current study is to reveal the effect of global linear transformations (shearing, horizontal stretching, and vertical stretching) on the recognition of familiar faces (e.g., a mother's face) in 6- to 7-month-old infants. In this experiment, we applied the global linear transformations to both the infants’ own mother's face and to a stranger's face, and we tested infants’ preference between these faces. We found that only 7-month-old infants maintained preference for their own mother's face during the presentation of vertical stretching, while the preference for the mother's face disappeared during the presentation of shearing or horizontal stretching. These findings suggest that 7-month-old infants might not recognize faces based on calculating the absolute distance between facial features, and that the vertical dimension of facial features might be more related to infants’ face recognition rather than the horizontal dimension.     

Identity-specific face adaptation effects: evidence for abstractive face representations

The effects of selective adaptation on familiar face perception were examined. After prolonged exposure to photographs of a celebrity, participants saw a series of ambiguous morphs that were varying mixtures between the face of that person and a different celebrity. Participants judged fewer of the morphs to resemble the celebrity to which they had been adapted, implying that they were now less sensitive to that particular face. Similar results were obtained when the adapting faces were highly dissimilar in viewpoint to the test morphs; when they were presented upside-down; or when they were vertically stretched to three times their normal height. These effects rule out explanations of adaptation effects solely in terms of low-level image-based adaptation. Instead they are consistent with the idea that relatively viewpoint-independent, person-specific adaptation occurred, at the level of either the “Face Recognition Units” or “Person Identity Nodes” in Burton, Bruce and Johnston’s (1990) model of face recognition.     

Stereotype priming in face recognition: interactions between semantic and visual information in face encoding

The accuracy with which previously unfamiliar faces are recognised is increased by the presentation of a stereotype-congruent occupation label [Klatzky, R. L., Martin, G. L., & Kane, R. A. (1982a). Semantic interpretation effects on memory for faces. Memory & Cognition, 10, 195-206; Klatzky, R. L., Martin, G. L., & Kane, R. A. (1982b). Influence of social-category activation on processing of visual information. Social Cognition, 1, 95-109]. For example, providing the label 'criminal' both during encoding and test improves recognition for previously unfamiliar faces that look like the stereotypical criminal. Experiments 1 and 2 both replicate this effect and show that the label exerts its influence during the encoding of stereotypical faces and has little influence at test. These findings indicate that semantic information that is congruent with novel stereotypical faces facilitates their encoding.     

Long-term effects of covert face recognition

Covert face recognition has previously been thought to produce only very short-lasting effects. In this study we demonstrate that manipulating subjects' attentional load affects explicit, but not implicit memory for faces, and that implicit effects can persist over much longer intervals than is normally reported. Subjects performed letter-string tasks of high vs. low perceptual load (Lavie, N. (1995). Perceptual load as a necessary condition for selective attention. Journal of Experimental Psychology: Human Perception and Perfomance. 21, 451-468.), while ignoring task-irrelevant celebrity faces. Memory for the faces was then assessed using (a) a surprise recognition test for the celebrities' names, and (b) repetition priming in a face familiarity task. The load manipulation strongly influenced explicit recognition memory, but had no effect on repetition priming from the same items. Moreover, faces from the high load condition produced the same amount of priming whether they were explicitly remembered or not. This result resolves a long-standing anomaly in the face recognition literature, and is discussed in relation to covert processing in prosopagnosia.     

Word recognition and face recognition following posterior cerebral artery stroke: Overlapping networks and selective contributions

It has been suggested that word- and face recognition rely on a common cerebral network for perceptual processing. To further explore this hypothesis, face- and word recognition were assessed in seven patients with focal lesions after stroke of the posterior cerebral artery in either hemisphere. The aim was to investigate if problems in face- and word recognition would co-occur, if testing was sensitive and patients were not pre-selected based on behavioural symptoms. Patients and matched controls were given a reading test and a delayed matching paradigm with faces, objects, and words. Interestingly, all patients with word recognition difficulties had problems in face recognition, independently of the affected hemisphere, supporting the existence of a bilateral perceptual network for faces and words. However, two patients showed selective face recognition problems after unilateral damage to either hemisphere, suggesting that parts of this network may be more critical for face than for word recognition.     

Multiple perceptual strategies used by macaque monkeys for face recognition

Successful integration of individuals in macaque societies suggests that monkeys use fast and efficient perceptual mechanisms to discriminate between conspecifics. Humans and great apes use primarily holistic and configural, but also feature-based, processing for face recognition. The relative contribution of these processes to face recognition in monkeys is not known. We measured face recognition in three monkeys performing a visual paired comparison task. Monkey and humans faces were (1) axially rotated, (2) inverted, (3) high-pass filtered, and (4) low-pass filtered to isolate different face processing strategies. The amount of time spent looking at the eyes, mouth, and other facial features was compared across monkey and human faces for each type of stimulus manipulation. For all monkeys, face recognition, expressed as novelty preference, was intact for monkey faces that were axially rotated or spatially filtered and was supported in general by preferential looking at the eyes, but was impaired for inverted faces in two of the three monkeys. Axially rotated, upright human faces with a full range of spatial frequencies were also recognized, however, the distribution of time spent exploring each facial feature was significantly different compared to monkey faces. No novelty preference, and hence no inferred recognition, was observed for inverted or low-pass filtered human faces. High-pass filtered human faces were recognized, however, the looking pattern on facial features deviated from the pattern observed for monkey faces. Taken together these results indicate large differences in recognition success and in perceptual strategies used by monkeys to recognize humans versus conspecifics. Monkeys use both second-order configural and feature-based processing to recognize the faces of conspecifics, but they use primarily feature-based strategies to recognize human faces.     

Familiar and unfamiliar face recognition: A review

Since the 1970s there has been a continuing interest in how people recognise familiar faces (Bruce, 1979 Bruce, V. 1979. Searching for politicians: An information processing approach to face recognition. Quarterly Journal of Experimental Psychology, 31: 373–395. [Taylor & Francis Online], [Web of Science ®] , [Google Scholar]; Ellis, 1975 Ellis, H.D. 1975. Recognising faces. British Journal of Psychology, 66: 409–426. [Crossref], [PubMed], [Web of Science ®] , [Google Scholar]). This work has complemented investigations of how unfamiliar faces are processed and the findings from these two strands of research have given rise to accounts that propose qualitatively different forms of representation for familiar and unfamiliar faces. Evidence to suggest that we process familiar and unfamiliar faces in different ways is available from cognitive neuropsychology, brain scanning, and psychophysics. However, in this review we focus on the evidence, available from experimental investigations of how people recognise faces, for different types of representation existing for each type of face. Factors affecting recognition are evaluated in terms of how they apply to familiar and unfamiliar faces and categorised according to the nature of their impact. In the final section this evidence, along with recent advances in the field, is used to explore the way in which unfamiliar faces may become familiar and the factors that may be important for the development of familiar face representations.     

An own-race advantage for components as well as configurations in face recognition

The own-race advantage in face recognition has been hypothesized as being due to a superiority in the processing of configural information for own-race faces. Here we examined the contributions of both configural and component processing to the own-race advantage. We recruited 48 Caucasian participants in Australia and 48 Chinese participants in Hong Kong, and had them study Caucasian and Chinese faces. After study, they were shown old faces (along with distractors) that were either blurred (isolating configural processing), in which high spatial frequencies were removed from the intact faces, or scrambled (isolating component processing), in which the locations of all face components were rearranged. Participants performed better on the memory test for own-race faces in both the blurred (configural) and scrambled (component) conditions, showing an own-race advantage for both configural and component processing. These results suggest that the own-race advantage in face recognition is due to a general facilitation in different forms of face processing.     

The own-age face recognition bias in children and adults

Children recognize children's faces more accurately than adult faces, and adults recognize adult faces more accurately than children's faces (e.g., Anastasi & Rhodes, 2005 Anastasi, J. S. and Rhodes, M. G. 2005. An own-age bias in face recognition for children and older adults. Psychonomic Bulletin and Review, 12: 1043–1047. [Crossref], [PubMed], [Web of Science ®] , [Google Scholar]). This is the own-age bias. Research has shown that this bias is at least partially based on experience since trainee teachers show less of an own-age bias than do other adults (Harrison & Hole, 2009 Harrison, V. and Hole, G. J. 2009. Evidence for a contact-based explanation of the own-age bias in face recognition. Psychonomic Bulletin & Review, 16: 264–269. [Crossref], [PubMed], [Web of Science ®] , [Google Scholar]). The present research tested the own-age bias in three groups of children (age 4–6, 7–9, 10–12 years) and a group of adults in the recognition of three age groups of faces (age 7–9, 20–22, and 65–90 years). Results showed an own-age bias for 7- to 9-year-old children and adults. Specifically, children could recognize faces more accurately if they were less than two years different from their own age than if they were more than two years older or younger. These results are discussed in terms of short-term experience with faces creating biases, and this rapidly changes with age.     

Multi-template supervised descent method for face alignment

Supervised Descent Method (SDM) is a highly efficient and accurate approach for facial landmark locating and face alignment. In the training phase, it learns a sequence of descent directions to minimize the difference between the estimated shape and the ground truth in feature space. Then in the testing phase, it utilizes these descent directions to predict shape increment iteratively. However, when the facial expression or direction changes too much, the general SDM cannot obtain good performance due to the large variations between the initial shape and the target shape. In this paper, we propose a multi-template SDM (MtSDM) which can maintain high accuracy on training data and meanwhile improve the accuracy on testing data. Instead of only one model is constructed in the training phase, multiple different models are constructed by repeatedly inputting the images which have large variations on expressions or head poses. And in the testing phase, the distances between some specific landmarks are calculated to select an optimal model to update the point location. The experimental results show that our proposed method can improve the performance of traditional SDM and performs better than several existing state-of-the-art methods.     

Newborns' face recognition over changes in viewpoint

The study investigated the origins of the ability to recognize faces despite rotations in depth. Four experiments are reported that tested, using the habituation technique, whether 1-to-3-day-old infants are able to recognize the invariant aspects of a face over changes in viewpoint. Newborns failed to recognize facial perceptual invariances between profile and full-face poses (Experiment 1), and profile and 3/4 poses (Experiment 3). Conversely, newborns recognized the identity of a face through full-face and 3/4 poses (Experiment 2). This result cannot be explained as a consequence of newborns' inability to discriminate between the full-face and 3/4 points of view (Experiment 4). Overall, evidence was provided that newborns are able to derive a representation of an unfamiliar face that is resilient to a certain degree of rotation in depth, from full-face to 3/4 and vice versa.     

Perceptual expertise and the plasticity of other-race face recognition

In this paper, we argue that our ability to recognize own-race faces can be treated as a form of perceptual expertise. Similar to object experts (e.g., birdwatchers), people differentiate own-race faces at the subordinate level of categorization. In contrast, like novices, we tend to classify other-race faces at the basic level of race. We demonstrate that, as a form of perceptual expertise, other-race face recognition can be systematically taught in the lab through subordinate-level training. When participants learn to quickly and accurately differentiate other-race faces at the subordinate level of the individual, the individuating training transfers to improved recognition of untrained other-race faces, produces changes in event-related brain components, and reduces implicit racial bias. Subsequent work has shown that other-race learning can be optimized by directing participants to the diagnostic features of a racial group. The benefits of other-race training are fairly long-lived and are evident even 2 weeks after training. Collectively, the training studies demonstrate the plasticity of other-race face recognition. Rather than a process that is fixed by early developmental events, other-race face recognition is malleable and dynamic, continually being reshaped by the perceptual experiences of the observer.     

Spatio-temporal dynamics of face recognition in a flash: it’s in the eyes

We adapted the Bubbles procedure [Vis. Res. 41 (2001) 2261] to examine the effective use of information during the first 282 ms of face identification. Ten participants each viewed a total of 5100 faces sub‐sampled in space–time. We obtained a clear pattern of effective use of information: the eye on the left side of the image became diagnostic between 47 and 94 ms after the onset of the stimulus; after 94 ms, both eyes were used effectively. This preference for the eyes increased with practice, and was not solely due to the informativeness of the eyes for the task at hand. The bias for the eye on the left side of the image is explained in terms of hemispheric specialization. Although there were individual differences, most participants exhibited this pattern of effective use of information. An intriguing finding is that most participants displayed a clear sinusoidal modulation of effective use of attention through time with a frequency of about 10.6 Hz.     

Reducing the own-race bias in face recognition by attentional shift using fixation crosses preceding the lower half of a face

Hills and Lewis (2006) reduced White participants’ own-race bias (ORB) in face recognition by training them to attend to features critical for Black faces (lower portion of the face). Here, the ORB was investigated following a brief fixation cross either in the upper portion of the face (critical for White faces) or the lower portion of the face. Results showed that when the cross preceded the lower portion of the face, Black faces were recognized more accurately than White faces and vice versa when it preceded the upper portion of the face. A second experiment demonstrated that this effect disappears if the participants are forced to delay their responses by 4 s. These results suggest that an immediate attentional mechanism can attenuate the ORB when immediate attention is paid to diagnostic features but this can be overridden with increased time spent viewing faces.     

No strong evidence for lateralisation of word reading and face recognition deficits following posterior brain injury

Face recognition and word reading are thought to be mediated by relatively independent cognitive systems lateralised to the right and left hemispheres, respectively. In this case, we should expect a higher incidence of face recognition problems in patients with right hemisphere injury and a higher incidence of reading problems in patients with left hemisphere injury. We tested this hypothesis in a group of 31 patients with unilateral right or left hemisphere infarcts in the territory of the posterior cerebral arteries. In most domains tested (e.g., visual attention, object recognition, visuo-construction, motion perception), we found that both patient groups performed significantly worse than a matched control group. In particular, we found a significant number of face recognition deficits in patients with left hemisphere injury and a significant number of patients with word reading deficits following right hemisphere injury. This suggests that face recognition and word reading may be mediated by more bilaterally distributed neural systems than is commonly assumed.     

The role of higher level adaptive coding mechanisms in the development of face recognition

Developmental improvements in face identity recognition ability are widely documented, but the source of children’s immaturity in face recognition remains unclear. Differences in the way in which children and adults visually represent faces might underlie immaturities in face recognition. Recent evidence of a face identity aftereffect (FIAE), in which adaptation (exposure) to a particular identity causes a previously neutral face to take on the computationally opposite identity, suggests that adults code faces in an opponent fashion relative to an average face. One previous study showed comparable FIAEs in 8-year-olds and adults but did not demonstrate that adaptation was selective for high-level representations in both groups. Using a developmentally appropriate FIAE task, we investigated whether children show adult-like adaptation for facial identity when adapting and test images differ in size. Both age groups showed an equivalent FIAE, suggesting that qualitative changes in the use of higher level adaptive coding mechanisms do not drive the developmental improvements in face recognition ability, at least from 8 years of age.