Sex differences in face gender recognition in humans

Brain Research Bulletin 63 (2004) 443–449 Sex differences in face gender recognition in humans Alessandro Cellerino a,b,∗ , Davide Borghetti c , Ferdinando Sartucci b,c a Institute of Neurophysiology of the Italian National Research Council (CNR) PISA, CNR, Via Giuseppe Moruzzi 1, 56100 Pisa, Italy b Scuola Normale Superiore, Pisa, Italy c Department of Neuroscience, Clinical Neurology, University of Pisa, Pisa, Italy Received 5 November 2003; received in revised form 11 March 2004; accepted 11 March 2004 Available online 15 June 2004 Abstract Human faces are ecologically-salient stimuli. Face sex is particularly relevant for human interactions and face gender recognition is an extremely efficient cognitive process that is acquired early during childhood. To measure the minimum information required for correct gender classification, we have used a pixelation filter and reduced frontal pictures (28,672 pixels) of male and female faces to 7168, 1792, 448 and 112 pixels. We then addressed the following questions: Is gender recognition of male and female faces equally efficient? Are male and female subjects equally efficient at recognising face gender? We found a striking difference in categorisation of male and female faces. Categorisation of female faces reduced to 1792 pixels is at chance level whereas categorisation of male faces is above chance even for 112 pixel images. In addition, the same difference in the efficiency of categorisation of male and female faces was detected using a Gaussian noise filter. A clear sex difference in the efficiency of face gender categorisation was detected as well. Female subject were more efficient in recognising female faces. These results indicate that recognition of male and female faces are different cognitive processes and that in general females are more efficient in this cognitive task. © 2004 Elsevier Inc. All rights reserved. Keywords: Face perception; Gender differences; Spatial filtration; Evolutionary psychology 1. Introduction The human face presents a clear sexual dimorphism [7,10,11,18,19]. Face gender recognition is an extremely efficient and fast cognitive process [6]. Even when images are cropped to remove all cultural cues to gender such as hairstyle and make-up, gender classification is correct in almost 100% of the cases in adult subjects [6], whereas 7-years-old children already reach 80% accuracy in the same task [21]. These data clearly indicate that biological cues in facial anatomy are sufficient for a very efficient gender recognition and this ability is acquired early during childhood. Male and female faces differ both for shape and texture and both shape and texture cues are used for face gender recognition. In frontal views, texture is more salient than ∗ Corresponding author. Tel.: +39 050 3153198. E-mail address: [email protected] (A. Cellerino). 0361-9230/$ – see front matter © 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.brainresbull.2004.03.010 shape for gender classification whereas the salience of shape becomes greater in lateral view [5,12]. Isolated facial parts can be used for face gender classification and the eye region has the highest load in judging gender, followed by the face outline [3,16,22]. Some global aspects of face configuration, however, are also important for face gender classification [16,17]. The present study was undertaken to answer these two basic questions: 1. Is gender recognition equally efficient for male and female faces? 2. Is there an interaction between gender of the observer and gender of the target face? Since ceiling effects are prominent in face gender recognition, to address these questions it is necessary to increase the difficulty of the gender classification task. Masking by spatial filtration is a widely-used technique in psychophysics. We have used two different and complementary spatial filtration techniques to mask pictures of male and female faces 444 A. Cellerino et al. / Brain Research Bulletin 63 (2004) 443–449 and study the effect of spatial filtration on gender recognition. In one case, a pixelation filter was used. This type of filtration greatly disturbs shape information but spares information concerning average colour distribution in the image. The second filter we have used is a Gaussian noise filter: it destroys information concerning average colour composition of the image leaving detection of high-contrast edges relatively spared. We found that both modalities of spatial filtration affect recognition of female faces more than recognition of male faces. In addition, male subject were more efficient in recognising male faces whereas female subjects were more efficient in recognising female faces regardless of the spatial filtration method used. 2. Materials and methods 2.1. Subjects One hundred twenty-one healthy observers (56 males, age span 17–63, 30 ± 5 years and 65 females, age span 19–46, 28 ± 5 years) voluntarily participated at Experiment 1. Fifty-five healthy observers (33 males, age span 17–63, 29 ± 6 years and 22 females, age span 20–46, 29 ± 5 years) participated in Experiment 2. Fifty-two healthy observers: (24 males, age span 19–56, 32 ± 5 years and 28 females, age span 19–45, mean 27 ± 5 years) participated in Experiment 3. The subjects were previously examined to exclude any visual problem or were corrected to their best acuity. All subjects signed an informed consent before participating to the test. 2.2. Experimental design and protocol 2.2.1. Experiment 1 Fifty frontal pictures of human male and female faces in neutral expression (25 M and 25 F; 260 × 377 pixels) were selected from a free database available on the web (Martinez A.M. & Benavente R., 1998; http://rvl1. ecn.purdue.edu/∼aleix/aleix face DB.html). Only pictures of individuals that appeared to be in the 20–30 years age range were selected. Males faces were all clean shaven and female faces were not wearing make-up. Each photo was included in a white oval frame using Adobe Photoshop 6.0® to remove hairs as well as facial outline. Spatial filtering was applied, using four different levels of pixelation (4, 8, 16 and 32 pixels; see Fig. 1A and B). Original pictures were composed of 28,672 pixels, the progressive filtration reduced the images to 7168, 1792, 448 and 112 pixels, respectively. The dimension of the filtered images on the computer screen were the same for all images. The whole set of images was 250, 50 not filtered and 200 filtered. These images were shown at 72 dpi on a 14′′ LCD laptop monitor at a viewing distance of 40 cm using a custom software. The software presented a randomised sequence of 50 faces that Fig. 1. Effect of spatial filtering on male and female faces: the two images on the left represents a male (A) and a female (B) face with a superimposed oval frame to remove facial outline and other features like necklaces or earrings. On the right, the same images were elaborated with a 32-pixelation spatial filter. Please note that the masking effect of the pixelation filter is dependent upon the size of the image, this can be easily appreciated by moving away the printed page or zooming out the pdf file on the computer screen. To perceive the same masking effect that was presented to the experimental subjects images should be seen at the full magnification of a 14′′ computer screen. were extracted from the set of 250 pictures and recorded for each subject the percentage of correct guesses for the unfiltered picture and for each of the four filtration levels. 2.2.2. Experiment 2 A further elaboration of images was then performed: we constructed ambiguous stimuli by alternating stripes of male and female faces (Fig. 2A). When filtered with the 32-pixelation filter, the stripes become invisible (Fig. 2B). An oval frame was also included. These images replaced the original male and female faces filtered with the 32 pixelation filter and an additional cohort of male observers was subject to the same test described in Experiment 1. 2.2.3. Experiment 3 Original un-modified images, including only an oval frame, were elaborated with Photoshop 6.0® and a Gaussian noise Filter at different intensity (25, 50, 75 and 100%; A. Cellerino et al. / Brain Research Bulletin 63 (2004) 443–449 445 Fig. 2. Ambiguous stimuli. Stripes of male and female faces were alternated to create an ambiguous face (A). When a 32-pixelation filter was applied (B), the stripes become invisible. see Fig. 3A and B) was applied. As in Experiment 1, the same specific software was used to show the images on a 14′′ LCD laptop monitor at a viewing distance of 40 cm. 2.3. Statistical analysis All the data concerning the number of male or female faces correctly recognized for each grade of pixelation or noise were imported in Microsoft Excel® to perform statistical analysis. For each grade of pixelation (Experiment 1), mean percent of correct guesses for male faces for both male and female observers were compared with mean percent correct answers for female faces using Student’s t-test for paired data. The same procedure was applied to compare results obtained only from male or female observers. Regarding ambiguous stimuli (Experiment 2), the ratio between male and female categorisation was calculated using the z-test to demonstrate if it was statistically different from chance level. Data obtained from Experiment 3 were analyzed in the same fashion of those obtained from Experiment 1. Data obtained from male and female observers in Experiments 1 and 3 were then plotted separately for male and females faces. To perfom canonic variate analysis (CVA) [1], the data were imported in Statistica 5.0, Statsoft. CVA is a variation of discriminant analysis and is a statistical method that performs a rotation of the data points into a new set of coordinates that are a linear combination of the original coordinates (in our case, the percentage of correct guesses for each of the filtration level) so to maximise the separation of the two groups along the new set of axis. Then the Malahanobis distances (a generalised form of distance that weights the co-variance of data across the different dimensions) between the groups are computed and the probability that the data assigned to different groups are part of the same Fig. 3. Effect of Gaussian noise filtering on male and female faces: the two images on the left represents a male (A) and a female (B) face with a superimposed oval frame. On the right, the same images were elaborated with a 100% Gaussian noise filter. Also in this case the masking effect of the spatial filter is dependent upon the size of the image. To perceive the same masking effect that was presented to the experimental subjects images should be seen at the full magnification of a 14′′ computer screen. group is derived based on the Malahanobis distance of all members of the groups from the centroids of the two groups. For each grade of pixelation (Experiment 1), mean percent of correct guesses for male faces for both male and female observers were compared with mean percent correct answers for female faces using Student’s t-test for paired data. The same procedure was applied to compare results obtained only from male or female observers. Regarding ambiguous stimuli (Experiment 2), the ratio between male and female categorisation was calculated using the z-test to demonstrate if it was statistically different from chance level. Data obtained from Experiment 3 were analyzed in the same fashion of those obtained from Experiment 1. Data obtained from male and female observers in Experiments 1 and 3 were then plotted separately for male and females faces. 446 A. Cellerino et al. / Brain Research Bulletin 63 (2004) 443–449 Fig. 4. Experiment 1: results. The percentage of male and female faces correctly recognized for each grade of pixelation were plotted separately for both male and female observers, for males only and for females only. 3. Results 3.1. Experiment 1: effects of pixelation Face gender recognition of full-resolution pictures is extremely efficient [6]. In order to increase the difficulty of the recognition task we used a pixelation spatial filter to progressively reduce the amount of information available for gender classification (see Section 2). An example of the result of the 32 pixelation filter is shown in Fig. 1. When filtered by the 32 pixelation filter, faces are represented by only 112 pixels. Male (N = 56) and female (N = 65) subjects were presented with 50 random pictures of male and female faces that were included in a oval frame to remove hairs and face outline. Pictures were either un-manipulated digital images or filtered with 4, 8, 16 or 32 pixelation filter. Percent of correct answers was plotted separately for male and females faces (Fig. 4). When un-manipulated faces were presented, both male and female faces were categorised correctly in about 95% of the cases, despite the elimination of face outline that is a strong cue to face gender [3,22]. However, the pixelation filter disturbed categorisation of female faces more effectively than it did for male faces. When the correct categorisation of faces filtered with 4 pixel filter were compared, categorisation of male faces was correct in 89% of the case whereas categorisation of female faces was correct in 82% of the cases (P = 0.006, paired t-test). Categorisation of female faces filtered with the 8-pixelation filter was at chance level, whereas categorisation of male faces was above chance even when the 32-pixelation filter was applied (P = 0.01, paired t-test; effect of sex of the face on the recognition efficiency tested by ANOVA, P < 0.0001). When the scores of male and female subjects are plotted separately, sex differences are apparent. Categorisation of faces filtered with 4-pixel filter did not differ for female subjects, (correct male faces = 88%, correct female faces = 87%, ns) whereas male subjects recognised male faces more efficiently than female faces (correct male faces = 90%, correct female faces = 77%; P = 0.002, paired t-test). The difference between male and female subjects in recognising female faces was statistically significant (unpaired t-test, P = 0.009). A difference in the efficiency of male and female subjects in recognising female faces was further supported by canonic variate analysis (see Section 2). Wilk’s λ-test detected a difference between the group of females viewing female faces and males viewing female faces with a significance of 0.03 (see Table 1). A trend of a difTable 1 Results of canonic variate analysis P-levels FvsM MvsM FvsF MvsF D2 distances FvsM MvsM FvsF MvsF FvsM MvsM FvsF MvsF 0 0.288 <0.0001 <0.0001 0.288 0 <0.0001 <0.0001 <0.0001 <0.0001 0 0,031 <0.0001 <0.0001 0.031 0 0 0.216 4.225 4.039 0.216 0 4.323 4.141 4.225 4.323 0 0.434 4.039 4.141 0.434 0 Results of a canonic variate analysis of face gender recognition. Coding of the groups is as follows. FvsF: female subjects viewing female faces; MsvF: male subjects viewing female faces; FvsM: female subjects viewing male faces; MsvM: male subjects viewing male faces. P-values are the result of an F-test based on the distribution of D2 distances withinhand between the groups. D2 is the distance of Malahanobis. The distance between a vector Xi and the mean vector Xm is defined as D2 = (Xi − Xm )Cx−1 (Xi − Xm ) where Cx−1 is the co-variance matrix of X. A. Cellerino et al. / Brain Research Bulletin 63 (2004) 443–449 447 Fig. 5. Experiment 3 results. The percentage of male and female faces correctly recognized for each percentage of Gaussian noise filter were plotted separately for both male and female observers, for males only and for females only. ferences in the categorisation of male faces was present as well. Categorisation of male faces filtered with the 32-pixel filter was only 60% correct and not statistically different from the chance level for female observers (P = 0.2, paired t-test), whereas male subject reached 72% accuracy level (P = 0.005, paired t-test). However, this difference was not significant neither in a unpaired t-test nor with canonic variate analysis (see Table 1). 3.2. Experiment 2: ambiguous stimuli Results obtained from Experiment 1 can be explained either by a higher efficiency of male faces categorisation or by a cognitive bias in categorising a face as male when the stimulus is ambiguous. To discriminate between these two possibility, we constructed ambiguous stimuli by alternating stripes of male and female faces (Fig. 2). When filtered with the 32-pixelation filter, the stripes become invisible (See Section 2). We then tested 35 additional male subjects using the same protocol but with all 32-pixelated faces substituted by ambiguous stimuli. The ratio between male and female categorisation for these ambiguous stimuli was nearly 0.5 and not statistically different from chance level (z-score = 0.84). 3.2.1. Experiment 3: Gaussian noise filter The presence of facial hair in fact makes the average colour of male lower faces darker even if the subjects are shaven. This low-resolution difference in colour would still be detectable after pixelation, as it is reproduced by the average pixel luminance, and could account for the higher performance of both sexes in the recognition of male faces. To discount this possibility, we filtered the same set of faces using Gaussian noise. In this second experiment, the power of the filter was calibrated so to be less perceptually destructive in order to pinpoint more clearly at differences in the recognition of female faces. Fifthy-four subjects, 24 males and 28 females, were presented with 50 random pictures of male and female faces as described in Experiment 1. Gaussian noise interfered more severely with the detection of female faces (Fig. 5, effect of sex of the face on the recognition efficiency tested by ANOVA: P < 0.001). In addition, when the performance of male and female observers were plotted separately, an interaction between sex of the observer and sex of the target face emerged: at maximum filter strength, the recognition of female faces was more affected by Gaussain noise in men observers than in female observers (unpaired t-test, P = 0.02; canonic variate analysis on the groups P = 0.03). 4. Discussion Face gender recognition is an extremely efficient cognitive process, nearing 100% correct guesses for frontal unkown pictures [6]. To test the existence of sex differences in face gender processing, we have increased the difficulty of face gender categorisation by using two different modes of spatial filtration: pixelation and Gaussian noise. Our study comes to two main conclusions: 1. Male faces are categorised more efficiently than female faces. 2. Subjects are more efficient in categorising same-sex faces. Previous studies have emphasised the high efficiency of face gender categorisation in adults and children [6,21], but a still open question is the minimal information needed to 448 A. Cellerino et al. / Brain Research Bulletin 63 (2004) 443–449 recognize face gender. An approach that has been undertaken is the categorisation using only facial parts as cues [4,22]. We have taken an opposite approach by employing spatial filtration: this procedure evenly degrades the quality of all facial parts without altering the general facial configuration. Using spatial filtration, we could increase the difficulty of face gender classification to find the conditions where recognition efficiency reaches chance level. We report that male observers can still correctly guess the gender of a male face in about half of the cases even when the face has been reduced to an image of only 112 pixels, i.e. an oval of 17 × 11 pixels, demonstrating that very little information indeed is required for correct male gender recognition. In addition, our experiments have demonstrated a striking difference in the recognition of male and female faces. Pixelation affects categorisation of female faces much more heavily than it does affect categorisation of male faces. Previous studies have reported a higher efficiency in the categorisation of male faces [21], but have interpreted this as the result of a cognitive bias. However, by using ambiguous stimuli, we have demonstrated that such a cognitive bias does not exist, at least in adult observers, and the differences we report reflects a true difference in the processing of male and female faces. In other words, less information is needed to correctly recognize a male face than it is needed to correctly recognise a female face. Since this difference is best revealed using pixelation filters, it can be concluded that very low spatial frequency information is sufficient to recognise male faces whereas the recognition of female faces requires higher spatial frequencies. It is interesting to note that a difference in categorisation for male and female faces is already observed in 7-years-old children [21]. Finally, we have found a sex difference in face gender recognition. In particular, female subjects are more efficient in recognising female faces. A trend was found for male subjects to be more efficient in recognising male faces than female subjects: this trend however was not statistically significant. This difference was observed regardless of the type of spatial filtration and has been already reported, in a more limited sample of Japanese subjects, using categorisation of facial parts [22]. It is not possible at present to provide a detailed explanation for this difference, but two alternative explanations could be indicated. This difference could be a mere consequence of differential socialisation. Since boys and girls tend to spend more time with individuals of the same-sex, they are also exposed for longer time to same-sex faces, hence the higher efficiency for recognition of same-sex faces. The alternative possibility is that the development and function of male and female brains is influenced by steroid hormones [14,15], and this influence extends also to face gender categorisation. In fact, sex differences in brain activation during face processing tasks have been reported [8,9,13,20]. More specifically, preliminary data from our laboratories have revealed sex differences in EEG response to the presentation of male and female faces [2] and ongoing experiment will test whether there is a neu- rophysiological correlate of sex differences in face gender recognition. Acknowledgements This work was supported by Scuola Normale Superiore grant SNS-03 and MIUR grant 2003062953. References [1] Anderson, An Introduction to Multivariate Statistical Analysis, Wiley/Interscience, New York, 2003. [2] D. Borghetti, A. Cellerino, F. Sartucci, Neurophysiological correlates of face gender recognition., Perception 32 (Suppl S) (2003) 175. [3] E. Brown, D.I. Perrett, What gives a face its gender? Perception 22 (1993) 829–840. [4] V. Bruce, A.M. Burton, E. Hanna, P. Healey, O. Mason, A. Coombes, R. Fright, A. Linney, Sex discrimination: how do we tell the difference between male and female faces? Perception 22 (1993) 131–152. [5] V. Bruce, S. Langton, The use of pigmentation and shading information in recognising the sex and identities of faces, Perception 23 (1994) 803–822. [6] V. Bruce, A. Young, In the eye of the beholder, Oxford University Press, Oxford, 1998. [7] A.M. Burton, V. Bruce, N. Dench, What’s the difference between men and women? Evidence from facial measurement, Perception 22 (1993) 153–176. [8] R.J. Erwin, R.C. Gur, R.E. Gur, B. Skolnick, M. Mawhinney-Hee, J. Smailis, Facial emotion discrimination: I. task construction and behavioral findings in normal subjects, Psychiatry Res. 42 (1992) 231–240. [9] D.E. Everhart, J.L. Shucard, T. Quatrin, D.W. Shucard, Sex-related differences in event-related potentials, face recognition, and facial affect processing in prepubertal children, Neuropsychology 15 (2001) 329–341. [10] M. Graw, A. Czarnetzki, H.T. Haffner, The form of the supraorbital margin as a criterion in identification of sex from the skull: investigations based on modern human skulls, Am. J. Phys. Anthropol. 108 (1999) 91–96. [11] R.J. Hennessy, A. Kinsella, J.L. Waddington, 3D laser surface scanning and geometric morphometric analysis of craniofacial shape as an index of cerebro-craniofacial morphogenesis: initial application to sexual dimorphism, Biol. Psychiatry 51 (2002) 507–514. [12] H. Hill, V. Bruce, S. Akamatsu, Perceiving the sex and race of faces: the role of shape and colour, Proc. R. Soc. London B Biol. Sci. 261 (1995) 367–373. [13] W.D. Killgore, Sex differences in identifying the facial affect of normal and mirror-reversed faces, Percept. Mot. Skills 91 (2000) 525–530. [14] D. Kimura, Sex, sexual orientation and sex hormones influence human cognitive function, Curr. Opin. Neurobiol. 6 (1996) 259– 263. [15] D. Kimura, Sex hormones influence human cognitive pattern, Neuroendocrinol. Lett. 23 (Suppl 4) (2002) 67–77. [16] A.J. O’Toole, K.A. Deffenbacher, D. Valentin, K. McKee, D. Huff, H. Abdi, The perception of face gender: the role of stimulus structure in recognition and classification, Membr. Cognit. 26 (1998) 146–160. [17] T. Roberts, V. Bruce, Feature saliency in judging the sex and familiarity of faces, Perception 17 (1988) 475–481. [18] A. Rosas, M. Bastir, Thin-plate spline analysis of allometry and sexual dimorphism in the human craniofacial complex, Am. J. Phys. Anthropol. 117 (2002) 236–245. A. Cellerino et al. / Brain Research Bulletin 63 (2004) 443–449 [19] M. Schmittbuhl, J.M. Le Minor, B. Allenbach, A. Schaaf, Shape of the orbital opening: individual characterization and analysis of variability in modern humans, Gorilla gorilla, and Pan troglodytes, Anat. Anz. 181 (1999) 299–307. [20] J.F. Thayer, B.H. Johnsen, Sex differences in judgement of facial affect: a multivariate analysis of recognition errors, Scand. J. Psychol. 41 (2000) 243–246. 449 [21] H.A. Wild, S.E. Barrett, M.J. Spence, A.J. O’Toole, Y.D. Cheng, J. Brooke, Recognition and sex categorization of adults’ and children’s faces: examining performance in the absence of sex-stereotyped cues, J. Exp. Child Psychol. 77 (2000) 269–291. [22] M.K. Yamaguchi, T. Hirukawa, S. Kanazawa, Judgment of gender through facial parts, Perception 24 (1995) 563–575.