A power analysis using G^*Power software version 3.01 (Faul et al., 2007) indicated that for an effect size of 0.2, at least 22 participants were required to achieve 80% power. In this study, 29 students (23 female students, aged 19.34 ± 1.21 years) completed the experiment, but one was excluded from the behavioral performance due to a data recording error. All participants were right-handed based on Edinburgh Handedness Test. All participants had normal or corrected-to-normal vision without color blindness or mental illness based on self-report. All study procedures were approved by the local Institutional Review Board of the School of Psychology, Shandong Normal University. Written informed consent was collected from all participants prior to the study initiation.

Each face stimulus was presented without hair or glasses and with a neutral expression, as generated by FaceGen Modeller 3.5 (Toronto, Canada, http://en.softonic.com/) and edited using Adobe Photoshop 7.0 (Adobe Systems, San Jose, CA). According to previous studies (Holmes et al., 2005; Dale and Arnell, 2014), HSF faces were processed in Photoshop using the high-pass filter tool to select only spatial frequencies higher than 6 cycles/degree or 30 cycles/image of visual angle (i.e., a radius of 1.5 pixels); LSF faces were processed in Photoshop using the Gaussian blur tool to select only spatial frequencies lower than 1.3 cycles/degree or 6.5 cycles/image of visual angle (i.e., a radius of 3 pixels); original images with broadband and no spatial frequency filtering were used for BSF faces (Figure 1A). A total of 24 faces were created (8 for each spatial frequency category), with the same grayscale, size, background, luminance, and other physical characteristics. The distractors were the same as those used in previous studies (Wang et al., 2015) and were the scrambled pictures, which were constructed by sectioning each picture into 272 rectangles (in a 7 by 9 matrix) and randomly reassigning the locations of these rectangles, without replacement.

All the stimuli were presented on a 17-in Lenovo monitor (resolution: 1,024 × 768; refresh rate: 60 Hz) using E-Prime 2.0 (Pittsburgh, PA, USA). The viewing distance was 60 cm and the stimulus size was 4° × 5.5° of the visual angle (113 × 156 pixels).

A delayed recognition task was selected for the study, and it allowed separate ERP investigations for encoding and retrieval stages (Morgan et al., 2008). Each trial started with the presentation of a central cross for 1,000–1,100 ms, followed by the encoding display (a 2 × 2 image array) for 1,000 ms. In the encoding display, one, two, or three different face stimuli were presented, corresponding to working memory load 1, load 2, and load 3, respectively, and the remaining stimuli were scrambled distractor images. The four different stimuli were presented randomly on the four corners of the fixation cross, and the eccentricity of the stimuli on the encoding display (measured as the distance between the center of each stimulus and the horizontal or vertical line of the central fixation cross) was 3.8° for stimuli in the four positions (Figure 1B). Participants were instructed to remember the identity of the faces and ignore the scrambled distractors. The stimuli presentation was followed by a 1,000–1,200 ms delay. A test face was then presented on the center of the screen for 1,000 ms (retrieval display), and participants were required to judge whether the test face appeared on the encoding display by pressing the keyboard (50% trials for “yes,” 50% trials for “no”). The intertrial interval was at random and ranged between 1 and 1.5 s.

A 3 (WM load: load 1, load 2, load 3) × 3 (face spatial frequency: low, high, broadband) within-subject design was adopted. Each condition had 120 trials, for a total of 1,080 trials.

EEG data were collected from 64 channels with an EEG recording system produced by the NeuroScan company. Vertical electrooculography (VEOG) was recorded with two electrodes in the upper and lower parts of the left orbital frontal region, and horizontal electrooculography (HEOG) was recorded with two electrodes placed 1.5 cm laterally in both eyes. The reference electrode was located between Cz and CPz. The resistance between all electrodes and the scalp was less than 5 kΩ. The band-pass filter range was 0.01–100 Hz, and the sampling rate was 500 Hz.

EEG data were analyzed using Letswave 7 software (https://letswave.cn; Mouraux and Iannetti, 2008) operating in Matlab R2013b (Mathworks, Natick, MA, USA). At both encoding and retrieval stages, the EEG analysis window was between −200 to 700 ms, and a baseline was acquired 200 ms before the appearance of stimuli. EEG data were band-pass filtered at 0.1–30 Hz, and the average of all electrodes was used as a reference. Eye artifacts were identified by independent components analysis and removed from all EEG electrode traces, and the rejection standard was ± 75 mV. Only trials with correct responses were analyzed for the retrieval stage.

According to previous studies (Wang et al., 2015, 2016, 2020), we analyzed P1 (75–145 ms), N170 (140–200 ms), and P2 components (200–270 ms) acquired from P3/P4, P5/P6, P7/P8, PO3/PO4, PO5/PO6, PO7/PO8, and O1/O2 electrodes. Based on previous studies (Morgan et al., 2008; Bauser et al., 2011), P3b (300–600 ms) was analyzed from PZ, POZ, and OZ electrodes to examine WM load manipulation. Besides, our stimulus number is low, which might make the faces familiar. To test the familiarity effect, the mean amplitude of N250r (230–320 ms) was analyzed from P7/P8, PO7/PO8, and O1/O2 electrodes (Neumann and Schweinberger, 2008). The latency (time from start to peak) and amplitude (baseline to peak) of each component were selected as dependent variables. A 3 (face spatial frequency: BSF, LSF, HSF) × 3 (WM load: load 1, load 2, load 3) × 2 (hemisphere: left, right) repeated-measures ANOVA was performed for P1, N170, P2, and N250r components. A 3 (WM load: load 1, load 2, load 3) × 3 (face spatial frequency: BSF, LSF, HSF) repeated-measures ANOVA was performed on reaction time and accuracy performance and the P3b component. When necessary, p values were corrected using the Greenhouse-Geisser method. The Bonferroni correction was applied to account for multiple comparisons and post-hoc analyses were performed to understand interaction effects.

There was a significant main effect of face spatial frequency on reaction time [F_{(2, 54)} = 3.51, p = 0.042, η2 p = 0.115]; however, post-hoc analyses did not detect any differences among the three SFs (LSF: 652 ms, HSF: 661 ms, BSF: 642 ms). A significant main effect of WM load was detected on reaction time [F_{(2, 54)} = 25.44, p < 0.001, η2 p = 0.485], and post-hoc analyses determined that the reaction time for load 1 was shorter than for load 2 and load 3 (614 vs. 667 vs. 674 ms, p < 0.001). The interaction between face spatial frequency and WM load on reaction time was not significant [F_{(4, 108)} = 1.54, p = 0.209].

There was a significant main effect of face spatial frequency on accuracy [F_{(2, 54)} = 39.13, p < 0.001, η2 p = 0.592], and post-hoc results revealed that accuracy was higher for BSF than for LSF and HSF conditions (79 vs. 74 vs. 71%, p < 0.05). The decomposition of a significant main effect of WM load on accuracy [F_{(2, 54)} = 321.86, p < 0.001, η2 p = 0.923] showed that accuracy decreased with increasing WM load (load 1: 88%, load 2: 75%, load 3: 62%, p < 0.001). Importantly, the interaction effect of face spatial frequency and WM load on accuracy was significant [F_{(4, 108)} = 3.22, p = 0.018, η2 p = 0.107]. The accuracy for detection of BSF faces was higher than for LSF and HSF faces under load 1 [91 vs. 87 vs. 85%, F_{(2, 54)} = 17.73, p < 0.001, η2 p = 0.396] and load 2 (80 vs. 73 vs. 70%, F_{(2, 54)} = 26.25, p < 0.001, η2 p = 0.493), and it was significantly different among BSF, LSF, and HSF conditions under load 3 (67 vs. 62 vs. 58%, F_{(2, 54)} = 34.27, p < 0.001, η2 p = 0.559). These results indicate that HSF appears to be less important than LSF under increasing WM load.

Figures 2, 3 show grand-average ERPs elicited by BSF, HSF, and LSF faces during encoding and retrieval stages, recorded from the temporal-occipital cortex. All stimuli evoked the canonical P1, N170, P2, P3b, and N250r components. Table 1 shows the results on P1, N170, and P2 amplitude and latency during the encoding and retrieval stages.

Previous studies reported that P1 was a robust component that was associated with holistic face processing and was more sensitive to LSF faces (Nakashima et al., 2008; Jeantet et al., 2019), which was in line with the assumption that configural information plays an essential role in early face processing as it could efficiently provide the whole structure of the face (Freire et al., 2000; James et al., 2001; Itier and Taylor, 2002, 2004). Conversely, others reported that P1 was more sensitive to BSF faces (Pourtois et al., 2005; Peters et al., 2013; Craddock et al., 2015). Consistent with these latter findings, our results showed that BSF faces elicited a larger and earlier P1 relative to HSF faces at the encoding stage. BSF faces were from the upright non-filtered faces, which are processed in a more configural way compared with inverted faces (Schwaninger et al., 2003; Kimchi and Amishav, 2010; Cousins et al., 2020). Interestingly, we found that LSF and HSF faces cannot be purely dissociated based on the early P1 component during the encoding stage, which might be reconciled by the assumption that HSF faces also inevitably contain configural information, less than LSF faces but enough to produce an effect similar to LSF faces during early face processing (Halit et al., 2006). The cutoff value of spatial frequency, which varied depending on studies, might be another important factor. LSF values were fewer than 8, 5, 4, or 3 cycles/image, and HSF values were above 35, 30, 24, 22, or 15 cycles/image (Jeantet et al., 2019). Our LSF and HSF faces might include mid-band spatial frequencies (MSF), which are important in face processing (for a review, see Jeantet et al., 2018). Thus, our results suggest that P1 was less sensitive to LSF and HSF faces during the encoding stage.

The impact of spatial frequencies on N170 modulation has been controversial. Our results replicated previous findings that N170 was more negative following the presentation of BSF or LSF faces relative to HSF faces during the encoding stage (Goffaux et al., 2003; Peters et al., 2013; Yao and Zhao, 2019), but are in contrast to other studies reporting a larger N170 in response to HSF faces (Nakashima et al., 2008; Jeantet et al., 2019) or reporting no association between N170 amplitude and face spatial frequency (Holmes et al., 2005). How can these discrepancies be reconciled? One answer may stem from selective attention to spatial location. Selective attention to objects in the periphery promotes LSF processing, and attention to foveal location promotes HSF processing (Shulman and Wilson, 1987; Carrasco et al., 2006). In this study, face stimuli were randomly presented in the four corners of the fixation cross during the encoding stage, directing attention to the periphery, hence the N170 amplitude was greater for LSF faces relative to HSF faces. Thus, the current N170 reflected the configural processing during the encoding stage.

Although many studies found that P2 is more sensitive to LSF pictures than HSF pictures (Mathes and Fahle, 2007; De Cesarei et al., 2013; Yang and Chan, 2015), P2 amplitude is also augmented by visual stimuli with broadband information (Hansen et al., 2011). In this study, P2 amplitude was more sensitive to BSF faces than to LSF and HSF faces during the encoding stage. According to previous research (Craddock et al., 2015), this might be related to the fact that BSF faces were upright non-filtered faces, which had higher spectral power and could drive stronger responses compared to faces with lower or higher spectral power. In addition, BSF faces contain not only configural and featural information but also MSF information (Halit et al., 2006). Previous research found that MSF information plays an important role in face processing (Collin et al., 2012; Jeantet et al., 2018) as it is intermixed with LSF and HSF information and can be processed in a similar way to the non-filtered BSF faces (Jeantet et al., 2019). Moreover, Parker and Costen (1999) demonstrated that recognition was more accurate and rapid for MSF faces than for LSF or HSF faces, which was also supported by later ERP research (Collin et al., 2012). These combined results led us to postulate that it might be the MSF information that made P2 highly sensitive to BSF faces during the encoding stage. Thus, our finding indicated that LSF and HSF faces cannot be dissociated from P2 amplitude during the encoding stages. But, the P2 latency showed that LSF faces were processed earlier than HSF faces.

In sum, for the encoding stage, LSF and HSF faces cannot be dissociated on the P1 component but can be dissociated on the N170 and P2 components, which were more sensitive to LSF faces than to HSF faces. The result is in line with the hypothesis that the N170 is a marker for face structural encoding and is linked to configural face processing (Eimer, 2011).

Consistent with our results at the encoding stage, the results at the retrieval stage showed that P1 was less sensitive to LSF and HSF faces and that N170 was more sensitive to LSF faces than HSF faces. However, the P2 was more positive for HSF faces relative to BSF and LSF faces during the retrieval stage, suggesting that P2 was more sensitive to face featural information. Although the sensitivity of P2 to featural processing has rarely been observed, several previous studies reported that faces with featural modifications elicited a larger P2 amplitude than faces with second-order relational modifications (Wang et al., 2016; Wang and Fu, 2018). Moreover, using Chinese and Western participants, Wang et al. (2020) found that P2 in Chinese participants was more sensitive to own-race faces with featural modifications. The reverse result was observed for Western participants, suggesting that cultural variation might account for this discrepancy.

It is worth emphasizing that the P2-enhancing effect of HSF faces was only observed at the retrieval stage and not at the encoding stage. It might be that faces in the encoding phase were processed in a parallel way not relying on in-depth visual analysis (cortical visual feedback; Mercure et al., 2008). Therefore, coarse facial information was more important during encoding, while during memory retrieval finer featural processing was adopted to match the encoded faces to the target at late stages (Bauser et al., 2011). Moreover, P2 could reflect the comparison between experimental face stimulus features and mental templates for task-related features (Potts, 2004); hence the facial features were of particular importance during the retrieval stage and evoked larger P2 waveforms. Taken together, only P2 reflected that encoding and retrieval stages were dissociated for face spatial frequencies.

In sum, for the retrieval stage, N170 was also more sensitive to LSF faces than to HSF faces, whereas P2 was more sensitive to HSF faces than to LSF faces. The processing priority of LSF faces during the retrieval stage is in line with previous neuronal findings that information processed in the LSF-sensitive magnocellular pathway has a faster cortical arrival than information processed in the HSF-sensitive parvocellular pathway (Laycock et al., 2007).

This study found that WM load modulated face SF processing, which was reflected in the behavioral results and ERP responses. Consistent with previous studies (Morgan et al., 2008; Bauser et al., 2011), the P3b amplitude decreased as WM load increased during both the encoding and retrieval stages. Moreover, at the retrieval stage, the N170 amplitude evoked by LSF and BSF faces did not differ but was larger than that for HSF faces, especially under the load 1 condition. These results indicated that BSF faces tended to be processed in a configural manner under a lower WM load, which agrees with previous findings that N170 reflects first-order relational processing (Maurer et al., 2002). However, our behavioral results showed higher accuracy in processing LSF faces relative to HSF faces under the high WM load condition, which might be related to the limited WM capacity for faces. Previous studies showed that the maintenance capacity for faces is approximately two (Towler et al., 2016), which might not have allowed participants to process the details and features of faces in the higher WM load conditions (load 3).

However, compared with the N170 component, we found a different result for the P2 component during WM retrieval. Specifically, the P2 amplitude evoked by HSF faces was larger than that for LSF faces, especially under the load 3 condition. Besides, BSF faces tended to be processed in a configural manner under the load 1 condition, as there was no difference between BSF and LSF faces. Conversely, they tended to be processed in a featural manner under the load 2 condition, as there was no P2 difference between BSF and HSF faces. BSF faces tended to be processed in both configural and featural manners under the load 3 condition, as there were no differences in P2 evoked by faces in any SF category. These results cannot be reconciled with the proposition from Morgan et al. (2008) that resources are limited during the retrieval stage by the increasing number of faces during the encoding stage. According to Morgan et al. (2008), attentional resources are sufficient under a low WM load, and fine processing of facial features should be adopted; under a high WM load, attentional resources are insufficient, and efficient global/configural processing should be adopted. However, we found the opposite results, and the mechanisms underlying this discrepancy need to be further investigated.

Notably, we also found a hemispheric advantage for N170 during the encoding and retrieval stages. In agreement with previous results (Rossion et al., 2003; Scott and Nelson, 2006), N170 was larger in the right hemisphere than in the left hemisphere during the encoding stage. Moreover, relative to HSF faces, the LSF faces elicited an earlier P2 under the load 3 condition during the encoding stage on the right hemisphere. This is in line with previous studies showing that configural processing occurs strongly in the right hemisphere (Scott and Nelson, 2006; Calvo and Beltran, 2014; Wang et al., 2016; Worley and Boles, 2016).

Our results might be confined to the familiar faces as the low stimuli number is likely to make the faces familiar in this study. We analyzed N250r, which is more sensitive to familiar faces than unfamiliar faces, to test the familiarity effect. Compared with HSF faces, LSF faces elicited more negative N250r in the present. This result provides evidence that familiar faces are more sensitive to configural information, even though there is an argument against this idea (Burton et al., 2015). Additionally, our data showed that the N250r amplitude decreased as the WM load increased during the encoding (from load 1 to load 2) and retrieval stages (from load 1 to load 3). The results were consistent with previous studies (Morgan et al., 2008), which employed six male faces. Further work is required to minimize the effect of familiarity. Another factor that cannot be neglected is the cutoff value of face spatial frequency. As we mentioned above, this value varied depending on studies, which might lead to different results. Compared with previous studies, our LSF (below 6.5 cycles/image) and HSF (above 30 cycles/image) faces might include the mid-band spatial frequencies [6.73–32 cycles/image in Collin et al. (2012); 5–15 cycles/image in Hsiao et al. (2005)]. Thus, this factor should be taken into consideration when comparing different studies.