We suggest greater aesthetic relevance, face for facial aesthetics, of the mobile and communicative parts for the female face and, conversely, of the rigid, structural, parts for the male one

Filtered beauty in Oslo and Tokyo: A spatial frequency analysis of facial attractiveness. Morten Øvervoll et al. PLOS, January 14, 2020. https://doi.org/10.1371/journal.pone.0227513

Abstract: Images of European female and male faces were digitally processed to generate spatial frequency (SF) filtered images containing only a narrow band of visual information within the Fourier spectrum. The original unfiltered images and four SF filtered images (low, medium-low, medium-high and high) were then paired in trials that kept constant SF band and face gender and participants made a forced-choice decision about the more attractive among the two faces. In this way, we aimed at identifying those specific SF bands where forced-choice preferences corresponded best to forced-choice judgements made when viewing the natural, broadband, facial images. We found that aesthetic preferences dissociated across SFs and face gender, but similarly for participants from Asia (Japan) and Europe (Norway). Specifically, preferences when viewing SF filtered images were best related to the preference with the broadband face images when viewing the highest filtering band for the female faces (about 48–77 cycles per face). In contrast, for the male faces, the medium-low SF band (about 11–19 cpf) related best to choices made with the natural facial images. Eye tracking provided converging evidence for the above, gender-related, SF dissociations. We suggest greater aesthetic relevance of the mobile and communicative parts for the female face and, conversely, of the rigid, structural, parts for the male face for facial aesthetics.

Discussion

A key question about what constitutes our sense of aesthetics is what kind of visual information within the stimulus underlies our judgements. Despite the spatial frequency structure of any visual stimulus is processed very early by the visual brain and several studies have addressed its role in the identification of facial identity and/or expression [81, 82], very few studies have specifically investigated the role of visual spatial frequency information in supporting our sense of facial aesthetics. That different face relevant types of visual information can be optimally channeled through different bands of spatial frequencies is well known for emotional expressions, but the possibility that a similar relationship occur for aesthetic cues has not been fully explored yet. It is very likely that other facial information, seemingly unrelated to visual spatial frequency, plays a relevant role in judgments of attractiveness (e.g., skin’s tone [83, 84]), but spatial frequency may play a role beyond the coding of facial shape. In particular, the optimal perception of several of surface and texture cues may be confined within specific bands of spatial information (e.g., the thin lines or creases revealing age or the colors of small parts like the irises). The appreciation of the colors (or discoloring) of small-width or thin facial parts (like the mouth lips) may also depend on high frequency information that may be smeared and significantly weakened in visibility at low spatial frequencies.
Hence, for the present study, we gathered evidence that forced-choice preferences when viewing specific SF bands of face images relate positively to preferences when viewing the corresponding broadband facial images. Our approach consisted in filtering spatial frequencies out of the natural face’s photo image (Fig 1). We then presented the obtained face images, containing a narrow band of SF information, in a “beauty contest” between same-sex face pairs (Fig 2). Although all of our photo images depicted faces of really existing European individuals, the participants of the present study belonged to different populations recruited in Europe and Asia (i.e., Norway and Japan).
The gender of the faces had a strong effect on which spatial frequencies were closest related to the same individuals’ decisions when performing the task with the unfiltered, natural looking, face pairs. That is, female faces related 80% of the time to choices made with the broadband faces, when viewing the highest of the four SF bands (Fig 3). In contrast, male faces related slightly above 80% of the time to choices made with broadband faces when viewing the second lowest of the four SF bands included in this study (Fig 3). We note that all of the SF bands related above 65% of time to choices made with broadband faces, indicating that all SF bands contribute to some extent with information relevant to aesthetics decision, although apparently in different doses. Thus, it would appear that medium-low spatial frequencies contains visual information that is most relevant for aesthetic decisions made about male faces, but the high spatial frequencies contain key information for decision about female faces. This dissociation is to our knowledge a novel finding, which could lead to identifying detailed gender-specific visual cues.
The oculomotor behavior provided converging evidence for the relevance of the medium-low spatial frequencies for male faces and high spatial frequencies for female faces. The attractive face in a pair was not only looked more in general than the unattractive, but gaze lingered the most over the attractive female face when the face pairs were shown with the highest (SF 4) filtering. Consistently, gaze dwelled the longest over the attractive male face when seeing the face pairs with the medium-low spatial (SF 2) filtering (Fig 4).
In addition, to get a sense of what information is contained and visible in the stimuli and within each band of SF information, we provided visualizations of how this content related to observers’ judgements. There was high similarity for the female faces between the two highest (SF 3 and 4) and the broadband faces within a small, bilateral, region (overlapping the eye pupil and the lower eyelid and infraorbital concavity but including the upper part of the zygomatic convexity; Fig 8). In addition, similarity between the filtered and unfiltered faces was higher for male models than female models for the area around the brows, nasal root, and eyes, and especially so for the (medium low) SF 2. Instead, for the medium high and high spatial frequencies (SF 3 and 4), the similarity was higher for female models than male models for the area around the lower eyelid and infraorbital region. Remarkably, the relative distributions of gaze when viewing these SF bands closely matched these similarity profiles.
Statistical analyses on the Lempel-Ziv complexity confirmed that the female faces contained significantly more information than the male faces in the above-described regions (Fig 9). However, information content was significantly high in the low SF bands only for the male faces; in particular, for the central eyebrow region, including the skull area immediately above (i.e., the supraorbital process or brow ridge) in SF 1, and the glabella of the nose and lowest nose region (including the nostrils) in SF 2.
Importantly, there was a striking dissociation between SFs for male and female faces in relation to the relevance of vertically oriented frequencies for attractiveness (normative) ratings. As visible in Fig 10, different spatial frequencies related to the ratings, revealing that for female faces, there was a significant positive correlation between attractiveness and amplitude of high vertical frequencies for female faces and at low vertical frequencies for male faces. Taking into consideration also the eye-tracking data, participants had a strong tendency to look at the faces along the whole axis of the nose (Fig 7), in particular in the European group, extending as low as the upper lip (philtum) and the Cupid’s bow at the center of the mouth, more so with increasing spatial frequency. This gaze behavior seems consistent with the preponderant role of the central, vertically oriented, features for attractiveness (normative) ratings.
Being the nose at the center of both the vertical and horizontal axes of the ‘face’ (Nota Bene: below the hairline, not the head), it is presumable that it constitutes an important element to focus gaze when evaluating facial proportions, the configuration and global harmony or symmetry of the face [58]. When spatial frequency is high, the volumetric aspect of the nose, relatively more relevant for the male face (Fig 11, right panel), becomes less visible. The nose is the most sexually dimorphic facial trait in its morphology, being on average disproportionally larger in volume in male than female faces [85–87]. While the visibility of the nose’s volume decreases that of its shape and symmetry increases with higher spatial frequencies and the latter features appear more relevant for judging the attractiveness of the female faces (Fig 11, left panel). Since gaze scanning (Fig 7) revealed a strong tendency to focus gaze at the root of the nose, or onto the central portion of the face that may correspond to the limiting size for efficient summation of configurational properties of upright (vertically oriented) face information in a single configurational face template [88]. The eyes, being paired features, horizontally centered together with the vertical prominence of the nose [89], may also convey essential information on a face’s proportions and symmetry, and more clearly so in higher spatial frequencies conditions.
It also seems of interest that the dispersion of gaze over the eyes, nose and mouth region differs in our European and Asian groups (Fig 7). The typical T-shaped focus pattern appears mainly with the European participants and increasingly so with higher spatial frequencies. In fact, the pattern of fixations is consistent with previous reports that Asians (i.e., Chinese) tend to look less at the eyes and distribute less their gaze over the face [90–92]. Especially within Japanese culture, a prolonged eye contact may be disrespectful and Japanese children are taught to look at others’ necks instead of the eyes [93, 94].
Perhaps the most remarkable dissociation between female and male features related to attractiveness, revealed by the present study’s Fourier approach, is between the two faces in Fig 11. These show graphic representations of the spatial frequencies that correlate positively with the stimuli’s normative attractiveness ratings (collected independently of the present eye-tracking study and only with Norwegian raters). A striking difference between the two genders’ images is that they show very different, little overlapping, SF components. Moreover, these SF components impressively overlap with the SF bands most relevant for forced choices, derived from the present eye-tracking study (Fig 3). For the male face (Fig 11, right panel), the attractiveness-correlated SF provide only a coarse visual resolution of the face, which however clearly conveys the depth or volumetric aspect of the head and face, with its overall size, extent of the face contour (the jaw and chin), and skull’s bone structure. These three-dimensional aspects of the male’s whole face or skull structure may be important in judging overall proportions. In contrast, the female face’s (in Fig 11, left panel) attractiveness-correlated SFs, not only show little overlap with the male’s, but they suggest that female attractiveness may be judged more on information carried by higher spatial frequencies. These may reveal local information about the surface of the face and of specific features at a level of detail that is optimal also for the task of individual person recognition and the communication of emotional signals.
In particular, internal features of the female face like the brow ridge, eyes, mouth, as well as the lower part of the face contour or chin, and their immediately surrounding facial surface regions, are clearly visible in the left panel image. We surmise that the high resolution of the above traits allows a more precise evaluation of the arrangements, spatial relations, or distance ratios between these features (e.g., the inter-ocular distance). There are several suggestions in the literature on facial beauty (also from anthropology, odontology, and aesthetic medical surgery) that our sense of face attractiveness may seek a “golden ratio” between facial traits like the eyes and mouth/teeth and the general proportions of the face ([95] but see [96]). We surmise that at HSF resolutions, information is optimal for spotting the presence of skin blemishes and the smoothness surface skin (i.e., cues of age or poor health) as well as details of the eye region affording the registering of subtle differences in eyelids’ and orbital region shape. If smooth skin is crucial for attractiveness in female faces and these properties of surface skin are best represented in high spatial frequencies, then amplitudes of higher frequencies should correlate with attractiveness ratings, since these frequencies make visible these aspects. We also note that the irises’ colors as well as the size of the pupils seem clearly delineated at such resolution. Instead, the colors of the irises would be smeared at LSF and, interestingly, previous research suggests that eye color may be more relevant when judging female than male faces for attractiveness [13]. Similarly, the highly mobile pupils may be particularly important for signaling social agreeableness, interest and attraction [8, 97]. We note that our behavioral and gaze results in the main experiment seem consistent with this ideas.
Moreover, the lower portion of the nose (nostrils) and the fullness of the lips (or vermilions) appear clearly visible within these attractiveness-correlated spatial frequencies and shape imperfections and coloring, luminance contrast between sides of the Cupid’s bow, may be very salient at this high resolution. Thus, female faces’ attractiveness-related SFs may reveal subtle deformations over the face surface, skin, and be related to the soft and malleable elements of the face, instead of its rigid skull structure. These highly mobile parts of the face like the mouth, eyes and eyebrows, all allow the display of subtle affiliative emotions [98], which may also play a key role when judging the attractiveness of an individual, even when just looking at static images [72].
In the male image in Fig 11 (right panel), a region around the ocular orbits, including the eyebrows and the bony area immediately above (i.e., the supraorbital process or ridge and glabella), as well as a region below the eyes and cheeks’ zygomas, appears well delineated in volumetry. Interestingly, the lower portions of the nose and of the mouth’s upper region play a role for male attractiveness, despite at such a coarse level of resolution the separations between the nostrils or lips are not resolved. Instead, the three-dimensional or volumetric aspects of the chin (in particular the protuberance of the mandible and its breadth) appear to be very salient. A possibility is that the coarse LSF prevalence in the image, by revealing the bony prominence of the brow ridge and of the jaw and chin, conveys effectively the attribute of masculinity inherent in the face [62, 99, 100]. In addition, a large face size characterizes masculinity as opposed to femininity [101]. However, several researchers have cautioned that masculinity may predict attractiveness relatively weakly compared to other fluctuating properties like skin color [102–104] or face and body symmetry [105, 106], which signal immunocompetence. Said and Todorov [18] found a gender-specific dissociation in the effects of shape (e.g., face width) or reflectance (e.g., lightness and color of skin). Increases towards masculinity in reflectance aspects of the male face increased attractiveness, but doing the same in shape aspects decreased it. We surmise that despite the coarse LSF male image (in Fig 11) both the reflectance of skin and of the brows are clearly visible. Interestingly, the reflectance dimensions with the strongest effects on female attractiveness involved the contrast around the eyes and the redness of the lips, which may be both best visible at higher SF.
Indeed, the HSF prominence in the image of the female face’ in Fig 11 yields a more detailed but somewhat less volumetric rendition (with slightly “embossed” features to use an art metaphor). What is visible appears related not only to highly mobile parts of the face that allow the display of subtle affiliative emotions but also to several cues associated with a sense of femininity [107, 108]. Sexual dimorphism correspond to different directions in morphometric space [108] and the female direction is associated with horizontal reduction of the chin, a forward movement of the gonion (jaw angle) and alveolar prognathism. In Fig 11, the male chin is clearly more visible than the female and appears larger in the morphed image.
The eye-tracking results confirmed that the beautiful faces are strong attractors of attention [109], since participants spent about 10% more time dwelling onto the attractive face in a pair (Fig 5) than on the relatively less attractive one. It has been shown that the attentional priority towards attractive faces can also occur unconsciously [110] and that a decision about a face’s level of attractiveness can be reached very rapidly (within 33 ms), and not very differently than when having unlimited time [111]. However, the present results are consistent with several previous studies showing that we typically spend extra time looking at faces considered attractive [112–114].
Finally, a previous study [38] used Fourier power spectrum analyses to describe the relation between spatial frequency and power of the radially averaged (1d) Fourier spectrum on a log-log scale. As the researchers point out, most natural (complex) images show a linear relationship and the relative strength or ‘power’ of fine detail information or coarse structure in an image can be, respectively, expressed linearly be the angle of the slopes in the power plots. Importantly, enhanced HSF information leads to shallow slopes, whereas enhanced LSF information leads to steep slopes. Given that pleasing natural scenes and artworks share a shallow power slope of -2 [115], the authors hypothesized that also faces approaching a Fourier power slope of -2 (i.e., with enhanced HSF information) would be considered more attractive than the same face, or others, differing from this value (e.g., steeper slopes between -3 and -4). Remarkably, when participants were given the opportunity to manually adjust the Fourier slope of the images on screen, they did choose a mean value of -2.6, which is a bit closer to that of pleasing natural scenes or artistic facial portraits. The effect was significantly larger for female faces, which also seems consistent with the present study’s findings of a bias for HSF information for female faces. A limitation of the Fourier slope approach is that it is informative about the relative distribution of frequency power, but not specific frequency bands. We surmise that, by presenting ranges of SF information separately, we are likely to reveal which information contained in the natural stimulus directly related to the aesthetic judgment about a face. In contrast, by strengthening or adding one type of visual information by distorting the natural image, one can reveal directional biases and explore the limits within which a face’s attractiveness can be enhanced [17].