We measured the effect of the correction of the natural aberrations of the eye by means of adaptive optics on the subject's performance on three different visual tasks: subjective sharpness assessment of natural images, familiar face recognition, and facial expression recognition. Images were presented through a dedicated psychophysical channel and viewed through an electromagnetic deformable mirror. Experiments were performed on 17 normal subjects. Ocular aberrations (astigmatism and higher order aberrations) were reduced on average from 0.366 ± 0.154 to 0.101 ± 0.055 *μ*m for a 5-mm pupil diameter. On average, subjects considered to be sharper 84 ± 14% of the images viewed under AO correction, and there was a significant correlation between the amount of corrected aberrations and the percentage of images that the subject considered sharper when observed under AO-corrected aberrations. In all eyes (except one), AO correction improved familiar face recognition, by a factor of ×1.13 ± 0.12 on average. However, AO correction did not improve systematically facial expression recognition.

^{2}.

*μ*m. In most cases, the residual was less than 0.15

*μ*m (RMS error correction of 71 ± 7%) except for one eye, which was that showing the highest amount of natural aberrations. A closed-loop correction (at a rate of 13 Hz) was typically achieved in 15 iterations. The state of the mirror that achieved this correction was saved and applied during the measurements when required in the psychophysical protocol. Psychophysical measurements were performed under static corrections of aberrations, as continuous dynamic correction would have involved continuous viewing of the spot test and discomfort to the subject (particularly given the relatively long duration of the test).

*p*(familiar ∣ familiar)) (true positive) and an unfamiliar face as familiar (

*p*(unfamiliar ∣ familiar)) (false positive). These probabilities were calculated as the fraction of faces identified with a certain confidence rating. For the first point of the ROC curve, the fraction of familiar faces receiving the rating 1 was plotted against the fraction of unfamiliar faces receiving the same rating. The following points of the curve were calculated as the cumulative fraction of the subsequent ratings. Therefore, the ROC curves represent the true-positive fraction versus false-positive fraction (Metz, 1978, 2006, 2007), i.e., the probability of correctly identifying a face as familiar versus the probability of incorrectly identifying an unfamiliar face as familiar. The same analysis is performed for happy/angry faces. The area under the ROC curves was measured with a trapezoidal method using the raw data. Perfect performance corresponds to an Au_ROC = 1. Inability to recognize any face or expression, i.e., completely arbitrary responses, will produce Au_ROC = 0.5. For some calculations and illustration purposes, we subtract the offset and defined Au_ROC′ = Au_ROC − 0.5. The analysis was performed separately for images viewed through natural aberrations and AO-corrected aberrations. The area under ROC curves was taken as a measure of recognition in each condition.

*SE*) of the difference between Au_ROC_noAO and Au_ROC_AO and follows the procedures of Hanley and McNeil (1982, 1983), which is similar to Wilcoxon (or Mann–Whitney). As the familiar face recognition experiment was conducted by presenting the images twice throughout the experiment (with both conditions, noAO and AO), the correlation between Au_ROC_noAO and Au_ROC_AO was taking into account, performing a correlated ROC analysis as described by Hanley and McNeil (1983). In the facial expression recognition experiment, we presented images only once (with either one or the other condition, noAO or AO). No correlation existed between Au_ROC_noAO and Au_ROC_AO, and therefore we applied a standard bivariate statistical analysis to calculate the Standard Error

*SE*.

*SE*of one ROC curve, we used Equation 1, from previous literature (Hanley & McNeil, 1982):

*n*

_{1}is the number of unfamiliar/angry faces (in the familiar face and facial expression recognition, respectively),

*n*

_{2}is the number of familiar/happy faces (in the familiar face and facial expression recognition, respectively), and Q1 and Q2 are calculated as Equations 2 and 3:

*SE*(Au_ROC_noAO − Au_ROC_AO)), annotated

*SE*(diff) in the Equation 4 for the facial expression recognition experiment (where different set of images were used in the two conditions AO, noAO) and Equation 5 for familiar face recognition where the same sets of images were used in both conditions and where Au_ROC_AO and Au_ROC_noAO are likely to be correlated.

*r*is a quantity representing the correlation introduced between the two areas by studying the same sample of faces (Hanley & McNeil, 1983).

*z*statistic using Equation 6:

*z*is above a critical level, we accept that the difference between the two ROC curves is significant. Typically, for a confidence level of 95%, the critical level is set at 1.96, which correspond to a type I error probability (

*p*-value, two tails) of 0.05 as a criterion for a significant difference.

*μ*m, with an average correction of 72 ± 7%. The estimated RMS excluding tilts, defocus and astigmatism was 0.244 ± 0.113

*μ*m. The average RMS values were calculated for 5-mm pupil diameters, except for S8 who had a 4.3-mm pupil.

*μ*m, with an average correction of 80 ± 10%, for 5-mm pupil.

*R*= −0.7,

*p*< 0.002 for RMS and

*R*= 0.5,

*p*< 0.05 for Strehl ratio).

*t*-test showed a statistical significant difference between the ROC curves under natural and AO-corrected aberrations across all subjects (

*p*= 0.0027).

*t*-test did not show a statistically significant difference between the areas under ROC curves, under natural and AO-corrected aberrations across subjects (

*p*-value = 0.8315).

*p*> 0.8).

*p*= 0.09) for familiar face recognition. For facial expression recognition, the gain with AO correction was always lower than 10% and did not show any correlation with the optical quality improvement in terms of RMS (AO/noAO).