Effect of low luminance on face recognition in adults with central and peripheral vision loss

To investigate the effect of low luminance on face recognition, specifically facial identity discrimination (FID) and facial expression recognition (FER), in adults with central vision loss (CVL) and peripheral vision loss (PVL) and to explore the association between clinical vision measures and low luminance FID and FER.

][7][8][9][10][11] Although most previous studies have investigated the relationship between vision impairment and face recognition under photopic conditions, people with vision impairment also report difficulties with face recognition under low luminance conditions, such as seeing faces in a dim restaurant. 1915][16][17][18] Furthermore, poorer discrimination of facial identity and poorer recognition of facial expressions under photopic conditions have been associated with greater disease severity and vision loss.With respect to AMD, most studies have shown that distance VA and CS are moderately to strongly associated with photopic recognition of faces and facial expressions, [5][6][7][8][9] with distance VA to a slightly greater extent than CS (r = −0.54 to −0.87 and r = 0.47 to 0.83 for VA and CS, respectively). As face recgnition is a central vision task requiring good VA, ocular conditions predominantly affecting peripheral vision (beyond 10° eccentricity), such as glaucoma, have received little attention until more recently.In two larger studies on glaucoma patients with good VA (inclusion criterion, 6/12 or better), it was found that, compared to age-similar adults with normal vision, patients with advanced bilateral Humphrey Field Analyzer (HFA) 24-2 visual field (VF) loss, significant HFA 10-2 VF loss or poor CS were less able to recognise faces.10,17 Therefore, these studies suggest that glaucoma patients with loss of central field sensitivity, despite good central VA, have reduced face recognition and that this seems to be in part mediated by reduced CS.It has been suggested that reduced central sensitivity in glaucoma might increase sensitivity to crowding, as is the case in the periphery of those with normal vision.15 Indeed, internal crowding by facial features has been demonstrated and suggested to occur in patients with AMD due to the need to use non-foveal vision.7,20 Given that conditions with predominantly central vision loss (CVL) and conditions with predominantly peripheral vision loss (PVL) (which often impact central vision to some extent) have been shown to decrease face recognition performance under photopic conditions, it would be useful to understand the effect of low luminance on face recognition in people with vision impairment from a variety of ocular conditions.
Most adults undertake many activities in low luminance conditions, with two studies reporting normally sighted adults being in conditions of <10 lux for approximately one-quarter of their waking hours. 21,22Recently, Dev et al. 23 found that community-dwelling older adults with and without AMD spent approximately one-third of their waking hours in low luminance conditions.Yet, adults with AMD report that they prefer brighter lighting and that faces and facial expressions are harder to recognise in low light levels. 1In another recent study that explored content for the development of a low luminance performancebased measure, people with vision impairment from various causes also reported difficulties with identifying faces and recognising facial expressions under low luminance conditions. 19Moreover, face recognition activities were ranked as highly important, second only to activities related to hazard detection and safety outdoors.What is not currently understood is the extent to which face recognition is reduced under low luminance compared to photopic conditions.
To date, the ability of people with vision impairment to recognise faces under low luminance conditions has been investigated in only one study.As part of an investigation comparing photopic face recognition performance of adults with AMD and those with normal vision, Bullimore et al. 5 tested the low luminance face recognition performance of a small sub-set of the sample (five participants with AMD and three with normal vision).Participants were required to correctly identify both the face and facial expression of single images.No formal analysis was undertaken due to the small sample; however, face recognition threshold was reduced for both AMD and normal vision control participants.Further research on a larger sample with various ocular conditions is required to understand the effect of vision impairment on face recognition under low luminance conditions.
The aim of this study was to investigate the effect of luminance on face recognition, specifically facial identity discrimination (FID) and facial expression recognition (FER), in adults with different types of vision impairment affecting central and peripheral vision, and to explore the associations between clinical vision measures and FID and FER performance under low luminance.In particular, we examined whether photopic vision measures that are easily and routinely assessed in the clinic could be used to adequately predict FID and FER, or whether it is necessary to assess low luminance vision measures.

Key points
• Low luminance significantly reduced face recognition, particularly for adults with central vision loss compared to those with peripheral vision loss and normal vision.• Poorer visual acuity and contrast sensitivity were associated with poorer facial identity discrimination and facial expression recognition under low luminance.• For clinical purposes, photopic visual acuity could be used as a good predictor of both photopic and low luminance face recognition.The study was approved by the QUT Human Research Ethics Committee and followed the tenets of the Declaration of Helsinki.Written informed consent was obtained prior to participation.

Clinical vision measures
Binocular photopic VA and low luminance VA (LLVA) were measured for all participants with distance habitual spectacle correction using an Early Treatment of Diabetic Retinopathy Study chart at 4 m. 24,25The luminance of the chart was 83.49 cd/m 2 (352.8lux, vertical at the chart) for the photopic condition and 0.72 cd/m 2 (2.8 lux) for the low luminance condition.As for all tests, illumination was achieved by overhead fluorescent lighting using a dimmer switch to produce mesopic conditions.VA and LLVA were recorded in logMAR, using a termination rule of four or more errors on a line and scored per letter. 26For participants who had severe vision loss and were unable to read any letters at 4 m, testing was conducted at 1 m, with appropriate adjustment to scoring.
Binocular peak CS and low luminance CS (LLCS) were measured for all participants with their habitual near spectacle correction using the Melbourne Edge Test (MET) at the recommended distance of 40 cm. 27The luminance of the MET chart was 47.51 cd/m 2 (452.4 lux) for the photopic condition and 0.49 cd/m 2 (5.5 lux) for the low luminance condition.CS and LLCS were recorded in decibels (dB), with a termination rule of two consecutive errors. 28or participants with PVL, VFs were assessed monocularly with the HFA 24-2 SITA standard program (Carl Zeiss Meditec, zeiss.com)and mean deviation (MD) used as a measure of generalised field loss.

Face recognition
Participants performed a FID task and a FER task.For both tasks, images of faces were chosen from the freely available validated Karolinska Directed Emotional Faces (KDEF) database. 29,30The KDEF database contains the faces of actors aged 20-30 years displaying seven different emotional expressions (neutral, happy, angry, afraid, disgusted, sad, surprised), photographed from five different profiles under standard photopic lighting conditions.For the current study, 36 colour images of faces in the straight-ahead position, with a closed mouth and no prominent facial hair, were selected from the database.[32][33][34][35] Faces were selected on the basis of having a high accuracy score for correct identification of the emotion being expressed (average accuracy score of all 36 images = 0.73 [SD: 0.21, range: 0.53-0.98],where 0 = never correctly identified and 1 = always correctly identified). 30ace images were downloaded as JPG files (562 × 762 pixels; 72 dpi; 24 bit-depth).

Facial identity discrimination task
Twenty-four images (12 female and 12 male) with a 'neutral' expression and having similar facial dimensions, features and background were selected.The original images were reduced by 60% to represent a life-size face at 3 m when viewed from 1 m (visual angle 3.5° height by 2.6° wide) based on previous reports for average face dimensions, 36,37 using an image editor (CorelDRAW Graphics Suite 2018 Version 20, corel draw.com).The decision was made to present the faces to be life-size at 3 m as this was considered a typical distance at which faces would be discriminated in the real world (e.g., identifying a familiar face among a group of people across the hall or street).For each trial, a set of three faces (two identical and one different) of the same sex were arranged side by side (Figure 1).Based on the similarity of facial dimensions and randomised location of the odd face (either left, middle or right), 15 sets of three faces were created from 72 possible combinations, three of which were used for practice, leaving 12 sets for testing FID.The sets of faces were printed in colour on A3 white paper and protected using matte laminate.The mean luminance of the sets of faces was 12.50 (SD: 2.20) cd/m 2 (323.9 lux) for the photopic condition and 0.17 (SD: 0.03) cd/m 2 (4.1 lux) for the low luminance condition.The 12 sets of test faces were presented in the same predetermined order at 1 m (see Table S1 for KDEF identification numbers of images and presentation order), and participants were asked to indicate the odd-one-out verbally (right, left or middle) and by pointing.

Facial expression recognition task
Twelve images (five female and seven male), four each with a neutral, happy or angry expression, and three practice images (two female and one male, one of each expression) were used for the FER task.Images were increased in size by 20% compared with the original to represent a life-size face at 1 m (visual angle 10.3° height × 7.7° wide), 36,37 using an image editor.A distance of 1 m was selected to represent a typical distance at which recognising facial expressions would be important for social interactions (e.g., interacting with a person in a conversation).Each image was printed and laminated as described above.Figure 2 shows examples of images.The mean (SD) luminance of the presented face images was 15.37 (2.62) cd/m 2 (322.9 lux) for the photopic condition and 0.19 (0.03) cd/m 2 (3.9 lux) for the low luminance condition.Single images were presented and participants asked to name the expression.A reference card with the options was provided to aid participants and minimise the effects of memory.The 12 test faces were presented in the same predetermined order (see Table S2 for KDEF identification numbers of images used and order).
Participants completed a short practice demonstration of the FID and FER tasks prior to testing.Both tasks were administered without any time restrictions and guessing was encouraged.Participants wore their habitual distance refractive correction.The order of photopic and low luminance testing was randomised with 10 min of adaptation time prior to each condition.Performance of each task was measured as the percentage of correct responses.

Data analysis
Statistical analyses were performed using IBM SPSS software for Windows version 27.0 (IBM, ibm.com), and p < 0.05 was used to indicate statistical significance.One-way analysis of variance was used to compare differences in demographic data and photopic vision measures between groups.As the data were potentially correlated (i.e., repeated measures) and accuracy of responses for the face recognition tasks was not normally distributed, the data were analysed using generalised estimating equations (GEEs), with a linear response and an exchangeable correlation matrix to allow for the correlation of the repeated measures within participants.Separate GEE models were run for FID and FER tasks, using luminance conditions (low and photopic) as within-subject factors, type of vision loss (CVL, PVL and control) as between-subject factors and interactions.
The correlations between age, clinical vision measures and the two face recognition tasks under low and photopic luminance separately were determined using Spearman's correlation coefficients.Stepwise linear regression analyses were then used to predict FID and FER accuracy under low luminance, using low luminance and photopic vision measures separately.

Participant characteristics
Thirty-three adults with CVL, 17 adults with PVL and 20 agesimilar controls participated in the study.Mean ages for the CVL, PVL and control groups were 75.6 (SD: 13.5) years, 70.0 (SD: 11.6) years and 71.6 (SD: 12.4) years, respectively.Fiftyfive per cent of the CVL group, 59% of the PVL group and 45% of the control group were female.The major cause of CVL was AMD (22/33; 66.7%); other causes were cone dystrophy, macular dystrophy, macular hole and optic neuropathy.For PVL the major cause was glaucoma (13/17; 76.5%); other causes were retinitis pigmentosa, optic neuritis and hemianopia.The differences between groups in mean VA and CS measures were statistically significant under both photopic and low luminance conditions (p < 0.001), with the CVL group having the worst VA and CS compared with the PVL and control groups (Table 1).

Facial identity discrimination versus face expression recognition
Comparing Figures 3 and 4, for the CVL group FID performance was significantly worse than FER performance under both low luminance and photopic conditions (p < 0.001).Similarly, for the PVL group, FID performance was significantly worse than FER performance under low luminance (p = 0.01) but not under photopic conditions.There were no significant differences for the control group under either luminance condition.p < 0.001; Table 2), with similar results for FID accuracy under photopic luminance.In the CVL subgroup analysis, low luminance and photopic measures of VA and CS were moderately to strongly correlated with FID accuracy under low luminance (ρ = 0.67-0.76,p < 0.001).Similarly, for the PVL group, low luminance and photopic measures of VA and CS were moderately to strongly correlated with FID accuracy (ρ = 0.61-0.77,p < 0.05).Also, HFA 24-2 MD in the better eye was moderately correlated with FID accuracy under low luminance (ρ = 0.54, p = 0.02; Table 2) but not with FID accuracy under photopic luminance.

Predictive models of face recognition under low luminance
Four separate stepwise linear regression models were run, one using low luminance vision measures (LLVA and LLCS) and one using photopic vision measures (VA and CS) to predict low luminance FID accuracy, and similarly, one using low luminance vision measures (LLVA and LLCS) and one using photopic vision measures (VA and CS) to predict low luminance FER accuracy.Prediction models of photopic FID and FER were not investigated given that the objective of the study was to explore the effect of low luminance.Vision measures entered into the regression models were based on the significant univariate correlations and all models were adjusted to account for the possible confounding effect of age.Given the relative consistency of univariate correlations across groups, models were run for all participants.All four models were significant predictors of both FID and FER accounting for approximately 61%-79% of the variance.The unstandardised regression coefficients are given in Table 4.Both low luminance and photopic measures of VA and CS were significant predictors of FID accuracy under low luminance after adjusting for age, explaining 79% and 75% of the variance, respectively.Only low luminance and photopic measures of VA were significant predictors of FER accuracy, explaining 65% and 61% of the variance, respectively.For both face recognition tasks, there was only a small increase in the variance explained by low luminance vision measures compared with photopic vision measures (4% for both FID and FER).

DISCUSSION
This is the first study to investigate the effect of low luminance on the accuracy of FID and FER in adults with CVL and PVL.Low luminance significantly reduced the accuracy of FID in adults with CVL and to a lesser extent in adults with PVL (mean reduction 20% and 8%, respectively).However, the accuracy of FER was significantly reduced only in adults with CVL (mean reduction 25%).Comparing the FID and FER tasks, for adults with CVL and with PVL, the FID task was more challenging and performance less accurate than the FER task.For both adults with CVL and with PVL, photopic and low luminance measures of VA and CS were moderately to strongly correlated with FID accuracy under low luminance.For adults with PVL, better eye HFA 24-2 MD was moderately correlated with FID accuracy under low luminance.Correlations were similar for FER.
Predictive models including all participants indicated that both low luminance and photopic measures of VA and CS were significant independent predictors of low luminance FID, with photopic VA and CS explaining 75% of the variance.However, only low luminance and photopic measures of VA were significant predictors of low luminance FER accuracy, with photopic VA explaining 61% of the variance and CS not making any additional contribution.Low luminance vision measures were only marginally more predictive of low luminance face recognition than photopic measures (4%).In this study, the negative effect of low luminance on face recognition was considerably greater for adults with CVL than for those with PVL, a finding consistent with a photopic study directly comparing the performance of AMD (predominantly causing CVL), glaucoma (predominantly causing PVL) and normal vision groups on a face memory test. 126][7][8][9] The reduced face recognition with CVL has been attributed not only to reduced spatial resolution but also to the need to adopt an extrafoveal retinal location to maximise visual functioning, 7,11,38 where VA and CS are reduced, 39,40 and crowding occurs (the inability to recognise objects in clutter), 41 all of which worsen with increasing retinal eccentricity and under low luminance conditions. 42More specifically, using caricatures of faces, Maritelli et al. 20 demonstrated that internal crowding of features within a face (self-crowding) becomes more prominent with increasing retinal eccentricity for observers with normal vision.Additionally, previous studies indicate that fixation is unstable in CVL, 43 and unlike adults with normal vision, fixation is more often directed towards external features (hair, chin and face outline) rather than internal features of faces (eyes, nose and mouth). 7,38,44As external features of faces are less detailed (contain more low spatial frequencies) and less cluttered than internal features, they are more resistant to poor spatial resolution and crowding. 7,45Therefore, it has been proposed that increased fixation towards external facial features might be used by some as strategy for extracting the most information as possible about a face rather than a simple consequence of CVL. 38Another strategy to mitigate the effects of spatial resolution and crowding is magnification.Although studies have shown that magnification can improve face recognition in adults with AMD, their performance does not reach that of individuals with normal vision. 8,11,13lthough not as poor as the low luminance FID performance of adults with CVL in this study, the performance of adults with PVL was worse than those with normal vision.Again, this is consistent with previous studies that have investigated face recognition under photopic conditions.7][48][49] Indeed, in the current study, the PVL group mean VA was approximately three lines worse and the CS 3 dB worse than the control group.Glen et al. 10 investigated the performance of glaucoma patients on a face memory test and found that those with advanced bilateral glaucoma performed poorly compared with patients with early and moderate glaucoma and age-similar people with normal vision.Moreover, glaucoma patients with significant central HFA 10-2 loss performed worse than those without loss of central sensitivity.Hirji et al. 18 confirmed this finding for recognising faces even in glaucoma patients with good VA, while Schafer et al. 15 showed results consistent with these studies for recognising facial expressions.Patients with glaucoma showed decreased performance and even those with good VA required a significantly shorter viewing distance (i.e., a larger angular subtense) than age-similar people with normal vision to recognise facial expressions (by 3.6 m). 15 Contrary to Schafer et al., 15 in this study, recognition of facial expressions was not significantly reduced for participants with PVL, a discrepancy likely due to differences in methodology.More specifically, it has been demonstrated that diffuse macular damage (mild generalised depression of sensitivity) impacts photopic face recognition more than focal macular damage (dense paracentral deficits), 17 and that diffuse glaucomatous macular damage is associated with self-reported visual difficulties under low luminance conditions. 50It has been suggested that crowding also plays a role in the diminished ability of those with glaucoma to recognise faces. 15egardless of luminance level, for participants with vision impairment in this study, FID was worse than FER performance.This finding is consistent with a study of participants with AMD under photopic conditions, 8 and could be due to the effective viewing distance being longer (and the visual angle of the faces smaller) for FID compared with the FER task, which were designed to represent real-world scenarios.However, Bullimore et al. 5 using a different methodology and varying the viewing distance, also found that participants with AMD performed worse when identifying faces compared with recognising facial expressions, such that the distance at which faces could be identified needed to be closer than the distance where expressions could be recognised.It is possible that this is because discriminating facial identity is a more complex task than recognising facial expressions, 51,52 requiring good central vision to detect, compare and integrate many facial features as well as interpret the face holistically. 20,52Indeed, the majority of studies on facial identity report that configural information (spatial interrelationships between face parts) plays a more prominent role than feature-based information (details of face parts). 51This seems to be less so for recognising facial expressions, for example recognising happiness largely depends on analysis of just the mouth; whereas recognising anger depends on an analysis of the eyes and mouth and their interrelationship. 51,53In normal central vision and in the periphery, happiness and surprise are better recognised than fear, anger and sadness. 54,55Also in the periphery, fear was better recognised than gender. 56ace recognition tasks such as FID and FER are not directly measured in clinics, and particularly not under low luminance; therefore, it is important to understand the relationship between clinical measures of vision, such as VA, CS and VF, and face recognition tasks.In the current study, both low luminance and photopic measures of VA and CS were moderately to strongly correlated with face recognition under low and photopic luminance, in both CVL and PVL, with little difference between low luminance and photopic findings.For PVL, better eye HFA 24-2 MD was also moderately correlated with face recognition under low luminance but not photopic conditions.][7][8][9] Although Hirji et al. 18 found similar univariate correlations with VA and CS as in this study, VA did not remain significant in multivariate analyses.Contrary to the current study that investigated FID and FER, Hirji et al. 18 found better eye HFA 24-2 MD was significantly correlated with performance on a photopic face memory test, but again this did not remain significant in multivariate analyses; instead, HFA 10-2 MD and CS remained significant in multivariate analyses.
In the current study, we were also interested in understanding low luminance face recognition and whether photopic vision measures that are easily and routinely assessed in the clinic could be used to adequately predict performance, or whether it is necessary to assess low luminance vision measures.In multivariate models, photopic VA and CS explained 75% of the variance in low luminance FID while photopic VA explained 61% of the variance in low luminance FER (with photopic CS not adding any further explanation).Using low luminance vision measures explained no more than an additional 4% of the variance in both low luminance FID and FER.This is the first study to quantify the effect of low luminance on face recognition in CVL and PVL.However, there are some limitations that should be considered.First, the group with PVL was relatively small and comprised heterogeneous ocular conditions.Furthermore, the face recognition tasks were a simulation using unfamiliar images of faces.Real faces and environments may provide additional cues and challenges to recognising faces and expressions.Although this study provides evidence of the detrimental effects of low luminance on face recognition in people with vision impairment, only one level of low luminance was evaluated and luminance levels in the real world can vary considerably and quite suddenly.Also, only distance letter acuity was measured, whereas in other studies a stronger correlation has been found between near word acuity and face recognition. 5,8Similarly, only HFA 24-2 was assessed and only for the PVL group.
In future, it would be useful to assess HFA 10-2 in both CVL and PVL, and to investigate whether specific points or areas within the central 10° are more important for face recognition than others.
In conclusion, this study demonstrated diminished face recognition under low luminance compared with photopic luminance conditions particularly for adults with CVL.Poorer VA and CS were strongly associated with worse face recognition under low luminance, with little difference between photopic and low luminance measures of VA and CS in predicting the accuracy of face recognition.The findings indicate that for practical purposes, clinical measures of VA are a good predictor of face recognition under both low luminance and photopic conditions.Overall, the current study provides a more comprehensive understanding of face recognition under low luminance in adults with CVL and PVL that will be useful for clinicians managing the challenges experienced by patients with low vision.[59][60]

F I G U R E 1
Examples of three sets of images used for the facial identity discrimination task, where participants were required to select the 'odd-one-out' [a.Karolinska Directed Emotional Faces (KDEF) identification numbers AM35NES and AM10NES; b.KDEF identification numbers AF19NES and AF28NE; c. KDEF identification numbers BM03NES and AM01NES].F I G U R E 2 Example of images used in the facial expression recognition task (a.neutral [Karolinska Directed Emotional Faces (KDEF) identification AF06NES], b. happy [KDEF identification AM14HAS] and c. angry [KDEF identification AM11ANS]).

F I G U R E 4
Photopic Characteristics of participants.
T A B L E 1Note: One-way analysis of variance.Abbreviations: CS, contrast sensitivity; HFA, humphrey field analyser; LLCS, low luminance contrast sensitivity; LLVA, low luminance visual acuity; MD, mean deviation; SD, standard deviation; VA, visual acuity.
Stepwise linear regression models of facial identity discrimination and facial expression recognition accuracy under low luminance for all participants (n = 70).
T A B L E 4