Making the incredible credible: Afterimages are modulated by contextual edges more than real stimuli

Abstract

Abstract

Abstract:

Abstract  We explored whether color afterimages and faint physical chromatic stimuli are processed equivalently by the visual system. Afterimage visibility in classic illusions appears to be particularly influenced by consistent contexts, while real stimulus versions of these illusions are absent in the literature. Using both a matching and a nulling paradigm, we present converging evidence that luminance edges enhance the perceived saturation of afterimages more than they do physical stimuli of similar appearance. We suggest that afterimages violate the response norms associated with real stimuli. This leads to the afterimage signal being ambiguous for the visual system, and thus more susceptible to modulation by contexts that increase or decrease the probability of the signal representing a real object. This could explain why afterimages are rarely experienced in everyday life, where they will be overruled by inconsistent context.

Introduction

Negative color afterimages occur when adaptation to a particular hue results in the subsequent illusory perception of its complementary hue. From this we could conclude that afterimages are simply negative images, or perceptions arising from a shift in relative activity of cells early in the visual system (e.g., retinal). According to this view, the independent adaptation of photoreceptor cells and subsequent shifts in the activity of opponent processes are mainly responsible for the generation of afterimages (Brindley, 1962; Craik, 1940; von Kries, 1970; Zaidi, Ennis, Cao, & Lee, 2012). The assumption follows that these signals are then processed by higher level visual areas equivalently to signals arising from any real stimulus. 

However, perceptually, afterimages do not always behave in the same manner as percepts of real objects. This could suggest that the visual system does not process afterimage signals equivalently to those arising from real stimuli. For example, afterimages are perceptually unstable: they can fade in and out of conscious perception (Comby, 1909; Wade, 1978). Moreover, they are highly dependent on context: in some situations we see them clearly, in others we do not, even given equivalent adaptation conditions (Daw, 1962). Classic afterimage illusions, such as the Spanish castle illusion (Sadowski, n.d.) and the recent illusion by van Lier, Vergeer, and Anstis (2009), demonstrate compelling effects of contextual modulators on afterimage visibility. In these illusions, afterimages embedded in a consistent context (usually a luminance edge or contour) are unequivocally visible, yet without such context they are much less visible, or even invisible. However, variants of these illusions displaying similar visibility modulation for real physical stimuli are notably absent in the literature. These observations led us to ask whether the representations of afterimages and real stimuli are substantially nonequivalent in the visual brain, and specifically, whether contextual cues (such as luminance contours) are particularly important for afterimage visibility. 

Previous psychophysical and electrophysiological research has demonstrated that luminance or chromatic edges are important in facilitating the perception of, and cellular responses to, physical chromatic stimuli to some extent (Friedman, Zhou, & von der Heydt, 2003). Perceptually, luminance contrasts (contours and pedestals) facilitate detection and discrimination of physical chromatic stimuli (Chaparro, Stromeyer, Kronauer, & Eskew, 1994; Cole, Stromeyer, & Kronauer, 1990; Eskew, Stromeyer, & Kronauer, 1994; Gowdy, Stromeyer, & Kronauer, 1999; Gur & Akri, 1992; R. Hilz & Cavonius, 1970; R. L. Hilz, Huppmann, & Cavonius, 1974; Montag, 1997; Mullen & Losada, 1994). In particular, a flashed suprathreshold luminance pedestal or contour (ring) facilitates detection of a coincident chromatic target (Chaparro et al., 1994; Cole et al., 1990; Eskew et al., 1994). Additionally, weak, blurry chromatic signals spread (i.e., fill in/fill out) until they reach a luminance edge (von der Heydt, Friedman, & Zhou, 2003). A demonstration of this process can be seen in the watercolor and Boynton illusions (Mollon, 1995; Pinna, Brelstaff, & Spillmann, 2001). At a physiological level, orientation selectivity and heightened responses to edges are common features of visual cortex cells (Friedman et al., 2003). There is also evidence of facilitatory interactions in the primate striate cortex (V1) between cells sensitive to luminance contrast and color (Horwitz, Chichilnisky, & Albright, 2005). 

It is possible that contexts, such as consistent luminance edges, are important cues to disambiguate real objects from variations in lighting. Most real objects possess clear luminance edges (Fine, MacLeod, & Boynton, 2003; Hansen & Gegenfurtner, 2009; Zhou & Mel, 2008), whereas this is not consistently the case for features less significant to awareness, such as variations in lighting (Kingdom, 2008), reflections, and afterimages. It is known that perceiving a colored surface as material rather than a light figment leads to an increase in its perceived saturation, suggesting that the visual system actively enhances the perceptions of objects (Bloj, Kersten, & Hurlbert, 1999). Although there are many possible means of distinguishing light from materials (see Kingdom, 2008, for a review), one possibility is that the visual system has learned to acknowledge, or evolved to enhance, faint chromatic signals when a luminance edge is present and disregard them when it is not. 

We were interested in whether a color afterimage that is matched to a real stimulus in hue, saturation, luminance, and edge blur and is viewed under fixation conditions is simply another example of a faint, chromatic stimulus. If this is the case, we would expect afterimages and real stimuli to be affected by context to a similar extent. Our general question is whether the brain treats afterimages and real stimuli as the same or whether there is an added uncertainty to afterimage signals that renders them more susceptible to contextual modulations. 

There may be a number of features of afterimage representations that distinguish them from responses to real stimuli, thus making afterimage signals more ambiguous. Importantly, we are not suggesting that afterimage signals and chromatic responses to real objects are not generated by the same cellular populations. Rather, we suggest that the nature of their responses—the temporal profile and distribution of signal strengths across different brain areas—may not be identical in all respects (see general discussion for elaboration). The visual system may then be faced with a dilemma over whether the afterimage signal should be perceived or suppressed. A surrounding context that is consistent with an ambiguous signal would raise the probability that it represents a real object, and thus raise the likelihood that it is perceived. If the signals underlying afterimages are by their nature more ambiguous than those for weak real stimuli, then we should expect that afterimage visibility will benefit more from a consistent context compared with a physical stimulus of similar appearance. 

We designed a series of complementary experiments to directly compare the enhancement effect of luminance edges on both color afterimages and suprathreshold physical stimuli. We used a simultaneous comparison task in which the chromatic contrast (saturation) of a physical stimulus was adjusted to match that of an afterimage (Experiment 1) and a nulling task in which the afterimage was nulled by a physical stimulus of complementary hue (Experiment 2). We then explored whether the effect of the contour on afterimages and physical stimuli was found for any surrounding edge, even if it was blurry, by substituting the contour for a sharp or blurred luminance pedestal (Experiment 3). We report converging evidence that sharp luminance edges (contours and pedestals) enhance afterimages to a greater extent than they do physical stimuli of similar appearance. 

Method and materials

Observers

For Experiment 1, eight observers (seven naive, one author; five males, three female) participated in both the afterimage and physical stimulus comparison tasks. A total of four observers (three naive, one author; three males, one female) participated in Experiments 2 and 3. All had normal color vision and normal or corrected-to-normal visual acuity. 

Apparatus

Stimuli were presented on a 21-inch Sony GDM-F520 Trinitron monitor (Sony) at 100 Hz, controlled by a Cambridge Research Systems (CRS) ViSaGe (Cambridge Research Systems) and a PC running Matlab (Manufacturer). Stimuli were viewed binocularly at a distance of 72 cm while the observer's head was maintained by a chin rest. Manual responses were made with a CRS CB6 button box. Eye movements were recorded by a CRS high-speed video eye tracker sampling at 250 Hz. 

Experiment 1: Afterimage and physical stimulus comparison task

Experiment 1 involved two stages. In the first stage, we measured the effect of luminance contours on the perceived saturation of afterimages by asking observers to compare the saturation of afterimages, framed or not by a contour, with physical comparison stimuli that varied in saturation. In the second stage, we measured the effect of the luminance contour on the perceived saturation of a physical reference stimulus that was similar to the afterimage in hue, luminance, and contrast. In other words, we substituted a real faint stimulus in place of the afterimage and based the properties of the physical stimulus on the afterimage matching results from stage one. 

Pilot studies

Before beginning Experiment 1, we conducted two pilot studies to ensure that our physical comparison patches were as similar in appearance to the afterimages as possible. One pilot calculated the appropriate amount of edge blur to introduce to the physical patches as a precaution against a chromatic edge overshadowing any effect of the luminance contour. The edge blur associated with afterimage percepts arises from fixational jitter during adaptation shifting the edges of the adaptation region. The second pilot perceptually equated the hue and luminance of physical patches with the afterimage, since in the main experiments only the chromatic contrast would be modulated. 

Stage 1 stimuli and procedure: Afterimage measurement

The chromaticity of the stimuli were originally calculated in MacLeod and Boynton color space (MacLeod & Boynton, 1979), but are reported in CIE chromaticity coordinates (x,y) and luminance in candelas per square meter (Y) (Smith & Guild, 1931) for convenience. The adapting stimuli were green (x = 0.252, y = 0.487, Y = 28.8) or pink (x = 0.303, y = 0.171, Y = 28.8) 3° diameter circles, presented 3° to the left or right of center. The physical comparison patches subtended approximately 3° (see explanation of edge blur below). One observer completed a pilot with a staircase to equate the hue and luminance of the comparison patches with the afterimages for a range of saturation levels (0, 10, 20, 30, and 40% of the adapting stimulus saturation). A second pilot was conducted to model the degree of edge blur predicted by fixational jitter during adaptation. One observer adapted to the same green/pink circles described above for 1.5 s across 360 trials, and eye movements were sampled every 4 ms. The retinal position (derived from the eye tracking data) of the stimulus at each 4-ms point during adaptation was then simulated, and these positions were translated into predicted stimulus edge blur during adaptation. This edge blur profile was then used to draw the comparison patches. 

Example trials are shown in Figure 1. All phases of the experiment were conducted on a grey background (x = 0.308, y = 0.316, Y = 28.8). During each trial of the main experiment, 8 observers fixated a black 0.15° diameter central dot and were adapted for 1.5 s on either the left or the right of fixation. Immediately following adaptation, one of five comparison patches was presented for 700 ms on the opposite side to adaptation. In half the trials, a 3° diameter contour (x = 0.308, y = 0.316, Y = 22.25) was presented opposite the comparison stimulus (to frame the afterimage). Observers were required to respond whether the left or right patch was “more saturated.” To reduce carryover adaptation, the comparison patches were followed by a 600-ms animated mask consisting of multiple 3° circles randomly changing position and chromaticity at 100 Hz. There were a total of 40 trial types within a 2 (adaptation hue) × 2 (contour presence) × 2 (test presentation side) × 5 (comparison patch saturation) within-subjects design. Observers received 10 repetitions of each trial type, totaling 400 trials, presented in a random order. Data were collapsed across test presentation side, resulting in 20 observations at each comparison patch level for each adapting color. Individual observer psychometric functions were fitted for each condition. From these we extracted the point of subjective equality (PSE), which represents the saturation level of the comparison patch perceived as equal to the afterimage. 

Figure 1

 

Example trials from Experiment 1. (a) Afterimage task, no contour condition. Observers fixate centrally and are then adapted to a pink or green circle on the left or right. One of five comparison stimuli is then presented on the opposite side to adaptation. Observers were required to indicate whether the left or right patch was “more saturated.” Trials ended with a mask to reduce carryover adaptation. (b) Afterimage task, contour condition. Procedure is identical but a contour is presented on the adaptation side during the test. (c) Physical stimuli task, no contour condition. Procedure is identical to afterimage task but without an adaptation phase and with a physical chromatic reference stimulus in place of the afterimage. (d) Physical stimuli task, contour condition.

Stage 2 stimuli and procedure: Physical stimulus measurement

The physical stimuli comparison task was identical to the afterimage task, but without the adaptation phase. The physical reference stimulus was a green or pink 3° diameter circular patch (with the same edge blur as the comparison patches) presented 3° to the left or right of center. The saturation value of the physical reference stimulus was set to the saturation level that matched the afterimage without a contour from stage 1. This was derived from the PSE between afterimage and physical stimulus in the no contour condition. These values were set individually for the green and pink afterimages and for each observer. The remaining design was the same as the afterimage stage. 

Experiments 2 and 3: Nulling task

In Experiment 2, we used a nulling task in which a physical stimulus of complementary hue was added to the afterimage until grey was perceived. The saturation of the physical stimulus (with and without a contour) was adjusted until the afterimage percept was nulled (the point of perceived grey; PPG). If afterimages are enhanced by luminance contours to a greater extent than physical stimuli of similar appearance, we would predict that more saturated physical stimuli are required to null the afterimages when both are framed by a contour. If the contour were to enhance both afterimage and physical nulling stimulus by the same amount, then these enhancements would cancel and the contour would produce no shift in the PPG (only an increase in slope might be expected due to enhanced color discrimination). We also repeated Experiment 2 with luminance afterimages to examine whether the effect was limited to color afterimages. All conditions of Experiment 2 were conducted in central vision while stimuli were presented peripherally in Experiment 1, which allowed us to investigate whether any effect of the contour differed with eccentricity. 

Experiment 3 followed the same nulling procedure as Experiment 2 but employed a luminance pedestal in place of a contour. Our hypothesis was that any luminance edge would produce afterimage enhancement, whether contour or pedestal. It was important to confirm this because although the afterimage illusions that motivated our research used luminance contours, most previous reports of chromatic facilitation of physical stimuli by luminance edges have used pedestals. Further, to examine the extent to which any effect of a pedestal was due to its edge rather than to the presence of a luminance component in the chromatic patch, we tested both sharp- and blurry-edged pedestals. Previous research has suggested that facilitation of physical chromatic stimuli by luminance signals is considerably more powerful with sharp, as opposed to graded, luminance differences. For example, discrimination of chromatic gratings is enhanced by superimposed square-wave luminance to a greater extent than sine-wave luminance gratings (Gowdy et al., 1999). 

General apparatus and stimuli description are the same as Experiment 1, unless stated otherwise. In Experiment 2, observers fixated a central dot and were adapted for 1.5 s to a central 3° diameter green or pink circle. Immediately following adaption, one of the seven nulling patches was presented for 700 ms. The nulling patches were drawn with edge blur profile determined in the pilot study for Experiment 1 and consisted of seven saturation vectors taken from the two adaptation hues: 10, 20, and 30% of pink and green, and one 0% (grey). The nulling patch was framed by a grey 3° diameter contour in half the trials. Observers were required to manually respond whether the patch appeared “pinkish” or “greenish” by pressing the appropriate response button. The nulling patches were followed by a 600-ms mask (described in Experiment 1). There were a total of 28 trial types within a 2 (adaption hue) × 2 (contour presence) × 7 (nulling patch hue/saturation) within-subjects design. Observers received 10 repetitions of each trial type, totaling 280 trials. 

For Experiment 3, the task, stimuli, and procedure were as described for Experiment 2. The single difference was that the contour was replaced by either by a sharp-edged luminance pedestal or a blurry-edged luminance pedestal that followed the blur profile used to draw the chromatic edges. There were a total of 42 trial types within a 2 (adaption hue) × 3 (pedestal type/ presence) × 7 (nulling patch hue/saturation) within-subjects design. Observers received 10 repetitions of each trial type, totaling 420 trials. Example trials for Experiments 2 and 3 are shown in Figure 2. 

Figure 2

 

Example trials from Experiment 2 and 3. (a) Experiment 2, no contour condition. Observers centrally fixate and are then adapted to a green or pink circle. One of seven nulling patches is then presented and observers are required to indicate whether the circle appears “pinkish” or “greenish.” Trial ends with a cycling mask to reduce carryover adaptation. (b) Experiment 2, contour condition. Procedure is identical but the nulling patch is surrounded by a luminance contour. (c) Experiment 3, sharp pedestal condition. Procedure is identical to Experiment 2, but a sharp pedestal is superimposed on the nulling patch. (d) Experiment 3, gradient pedestal condition. Same as above, but with a blurry pedestal.

Results

Experiment 1: Afterimage and physical stimulus comparison task

Figure 3 shows results from Experiment 1 and displays an interaction between contour and stimulus type, such that the difference between the perceived saturation in the contour and no contour conditions was larger for afterimages than for physical stimuli (F[1, 7] = 11.66, p < 0.01). No main effect of stimulus color was found and color did not interact with the stimulus type or contour presence. 

Figure 3

 

Results of Experiment 1. Mean point of perceived equality (units are percentage of adapting stimulus saturation) of afterimages (black) and real stimuli (white) in the contour and no contour conditions, across 8 observers. The significant interaction is clearly shown, whereby the contour enhanced the perceived saturation of afterimages more than physical stimuli of similar appearance. Error bars show, for each condition, the standard error of the differences from each participant's mean (i.e., they are derived from the portion of the variance that is relevant for within-subject tests by excluding the irrelevant main effect of subject). The mean for each point is derived from 20 data readings per observer.

These results indicate that even though the physical stimuli were similar to the afterimages in hue, luminance, degree of edge blur, and saturation, the luminance contour increased the perceived saturation of the afterimages significantly more than it increased the perceived saturation of the physical stimuli. 

Experiments 2 and 3: Nulling task

Figure 4a plots the PPG across four observers for Experiment 2, which represents the amount of physical stimulus saturation required to null the afterimage. For all observers, the PPG in the contour condition is shifted further away from physical grey than in the no contour condition (t[3] = 6.51, p < 0.01). This finding confirms that more physical stimulus saturation is required to null an afterimage that is framed by a contour. We found no consistent increase in slope in the contour condition, providing no evidence that the contour improves discrimination of the combined stimulus (the afterimage and the null). A steeper slope in the contour condition is predicted by previous research reporting that luminance edges improve detection and discrimination of real chromatic stimuli (Chaparro et al., 1994; Cole et al., 1990; Eskew et al., 1994; Gowdy et al., 1999; Gur & Akri, 1992; R. Hilz & Cavonius, 1970; R. L. Hilz et al., 1974; Montag, 1997; Mullen & Losada, 1994). These findings have also been replicated by us in a pilot study, results of which are shown in supplementary information. 

Figure 4

 

Results from Experiments 2 and 3. Grey lines represent individual observers; black is the mean across observers. Calculation of error bars is the same as described for Experiment 1. There are 20 data readings for each individual observer point. (a) Experiment 2. Point of perceived grey (PPG; in percentage of adapting stimulus saturation) in the no contour and contour conditions. Across all observers, more physical saturation is required to null the afterimage in the contour condition compared with the no contour condition. (b) Nulling procedure replicated with luminance afterimages. Individual data from four observers in the no contour, light contour, and dark contour conditions. The PPG, or the point at which nulling patch is judged perceptually equal in luminance to background, is plotted in log units from the background. Dark contours consistently shifted the null point compared with the no contour condition. Light contours shifted the PPG in a similar direction for three out of four observers. (c) Mean physical saturation required to null the afterimage (PPG; units are percentage of adapting stimulus saturation) across sharp, blurry, and no pedestal conditions across four observers. Figure illustrates that a sharp pedestal increases the perceived saturation of the afterimage compared with the no pedestal condition, whereas a blurry pedestal does not increase the saturation of afterimage above the no pedestal condition.

The contour effect on afterimages is not specific to chromatic afterimages. Figure 4b shows the results of a luminance afterimage nulling experiment (procedure identical to Experiment 2, but observers adapted to light, monochromatic patches and the resultant afterimages were framed by a light contour, dark contour, or no contour). As shown in Figure 4b, both light and dark contours tended to shift the PPG relative to the no contour condition, indicating that the afterimage enhancement exceeded that for the physical nulling patches. 

Figure 4c shows results from Experiment 3, in which sharp and blurred pedestals were presented instead of a contour. The mean shift in PPG across observers is greater in the sharp pedestal condition compared with the no pedestal and blurry pedestal conditions (F[2, 6] = 8.88, p < 0.05). Posthoc tests revealed that the perceived saturation of the afterimages in the sharp pedestal condition was significantly greater than that in both the blurry pedestal and no pedestal conditions (p = 0.041 and p = 0.023, respectively). In contrast, the blurry pedestal condition was not different to the no pedestal condition (p = 0.554). This suggests that the benefit in the sharp pedestal condition depends on the presence of a sharp edge rather than the mere presence of a luminance increase congruent with the color gradient. For all but one observer the psychometric slope in the sharp pedestal condition was steeper that the no pedestal or blurry conditions. This suggests that the discrimination of the combined stimulus (the nulling stimulus and the afterimage) was improved by a superimposed sharp pedestal, a finding consistent with previous findings for the enhancement of real chromatic stimuli. Further, the importance of edges for chromatic perception is demonstrated in examples of chromatic spreading and “filling in,” such as the Boynton and watercolor illusions (Mollon, 1995; Pinna et al., 2001). Why we observed improved discrimination in the sharp pedestal condition here, but not with the luminance contours in Experiment 2, is unknown. 

Taken together, the results of Experiments 1 through 3 suggest that the contextual modulators of afterimage visibility seen in recent compelling afterimage illusions (Sadowski, n.d.; van Lier et al., 2009) is not observed to the same degree for real, faint chromatic stimuli. 

Thus, although previous research has demonstrated that real chromatic stimuli are enhanced by luminance edges, our findings indicate that afterimages are enhanced by luminance edges more than physical patches. This finding implies that there may be something different about afterimage representations that results in particularly powerful modulations by contextual cues (in this case, luminance contours). 

Control experiments

We conducted two further experiments to control for other differences between the afterimages and physical stimuli that could have driven the different degrees of contour enhancement found in Experiments 1 through 3. 

Control 1 (Experiment 4): Were the edges of the physical stimuli blurry enough?

The larger contour effect for afterimages compared with real stimuli could be due to an underestimation of afterimage edge blur. If the edges of the physical stimuli are less blurry than those of the afterimages, we might expect to find less modulation by the contour because sharper chromatic edges may themselves contribute to the contour/edge effect. The edge blur profile used to draw the physical patches was based on eye movement jitter from 1 observer, and analysis of the jitter from other observers revealed some variance with this standard (Experiment 1: 3 lower, 5 higher than standard; Experiment 2: 3 higher, 1 lower than standard). Although no correlation was apparent between this jitter and the difference between the afterimage and physical stimulus contour effects, these correlational analyses had very low power. Therefore, we tested the effect of edge blur in a further comparison experiment, identical to the physical stimulus conditions described for Experiment 1. 

The edge blur of the reference stimulus was modulated from increased blur to sharp edged (methods in supplementary information) and presented with and without a luminance contour. The results showed that the contour effect did not increase as the edges of the patches became increasingly blurry (Figure 5). Specifically, there was no interaction between the contour effect and degree of blur when the reference stimuli were physically equally saturated across edge blur conditions (Figure 5a). This was also the case when the perceived saturation levels of the reference were equated across levels of edge blur (Figure 5b). Perceived saturation was equated to counteract changes in saturation due solely to edge blur, as we wanted to be sure that this did not interact with the contour effect. These findings indicate that even if the blurriness of our physical stimuli was underestimated for some observers in Experiments 1 through 3, this is unlikely to explain why we found increased contour modulation of the afterimages relative to physical stimuli because increasing edge blur does not lead to an associated increase in contour modulation. 

Figure 5

 

Results of Experiment 4. Perceived saturation of the reference from the physical saturation, across edge blur and contour conditions, when the patches were (a) physically equally saturated and (b) perceptually equally saturated. Physical saturation for the reference was 20. Both illustrate that although the contour increases the perceived saturation of the reference, it does not do so in a way that interacts with the blurriness of the reference. Error bars are calculated in the same manner as Experiment 1 (some are smaller than the marker size). There are 20 data readings for each point.

Control 2 (Experiment 5): Temporal order of contour and physical stimulus onsets

Another possible explanation for why afterimages were enhanced to a greater extent by edges compared with the physical stimuli is that adaptation signals are present in the visual system prior to the presentation of the contour or the physical stimuli. Indeed, as luminance signals tend to reach the visual cortex 10 to 30 ms before chromatic signals (Bompas & Sumner, 2008; Maunsell & Gibson, 1992; Nowak, Munk, Girard, & Bullier, 1995; Schmolesky et al., 1998) and the adapted color signal is present before this, the afterimage will benefit from any contour enhancement prior to the physical stimulus. In order to control for this difference, 3 observers repeated the afterimage and physical stimulus comparison task (Experiment 1). However, we varied the onset of the contour so that it was presented 0, 20, 40, or 60 ms after adaptation or physical stimulus onset. If the physical stimulus suffered from arriving in the cortex after the contour, we might predict its contour effect to increase for a contour delay of 20 to 40 ms. Against that prediction, results revealed that the contour effect remains fairly stable between contour onsets of 0 to 60 ms for both afterimages and physical stimuli, suggesting that contour onset does not notably modulate contour enhancement effects (Figure 6). Further, the contour effect is greater for afterimages than physical stimuli, thus replicating the results of Experiment 1. 

Figure 6

 

Results of Experiment 5. Perceived saturation of afterimages (black) and physical stimuli (white), with and without a contour, across four contour onset times. The no contour and 0-ms onset points are a replication of Experiment 1 and show that both afterimages and physical stimuli were perceived as equally saturated when not framed by a contour (as expected from the experimental design), but afterimages were perceived as more saturated when framed by a contour than were physical stimuli. However, the pattern is similar across different contour latencies, showing that contour onset time did not interact with the contour effect or the type of stimulus (physical or afterimage). Error bars are calculated in the same manner as Experiment 1. There are 20 data readings for each point.

General discussion

We investigated whether negative color afterimages are treated identically to real physical stimuli. We theorized that there are differences between the adaptation response that results in afterimage percepts and signals from physical stimuli, and that these differences may lead to a different degree of luminance edge modulation of afterimages compared with real stimuli. Our theory was motivated by noting the absence of physical stimulus versions of illusions demonstrating compelling contextual modulations of afterimage visibility. 

Afterimages are particularly enhanced by context

We report converging evidence from both a comparison and nulling paradigm that luminance edges (contours or pedestals) enhance the visibility of color afterimages to a larger extent than they do for physical stimuli of similar appearance. In the comparison experiment (Experiment 1) the physical stimuli were similar to the afterimages in appearance (hue, luminance, and saturation), yet were not enhanced by the contour to the same degree. The nulling paradigm (Experiments 2 and 3) revealed that more physical stimulus saturation was required to null the afterimages when framed by a luminance edge. This result excludes the possibility that the edge enhanced both the afterimage and the physical nulling patch equally, which could have resulted in an overall increase in discrimination when the contour was present, but not the observed shift in the null point. This is because the contour would enhance discrimination of the combined signal of the nulling patch and the afterimage. Our findings are not constrained to chromatic stimuli because we replicated the results of Experiment 2 using luminance afterimages. 

A control study (Experiment 4) revealed that the difference in contour enhancement between afterimages in physical stimuli was not due to an underestimation of the edge blur used to draw the physical patches (fixational jitter during adaptation will blur afterimage edges). We found that increasing or decreasing the blurriness of the physical patches did not lead to a decrease or increase in the size of the contour effect. A further control study (Experiment 5) revealed that our results do not arise from temporal differences in the onset of afterimage versus real signals. Contour enhancement was greater for the afterimages compared with physical stimuli, thus replicating the results of Experiment 1. However, delaying presentation of the contour (0–60 ms) after presentation of the physical stimulus did not impact the degree of contour modulation. 

Afterimages as ambiguous stimuli

Our finding that luminance edges enhance afterimages to a greater extent than they enhance physical stimuli suggests that afterimage representations are processed to some extent differently than signals arising from real-world objects. One possible reason for this effect is that afterimages are ambiguous, like other phenomena such as binocular rivalry and ambiguous figures. For example, the tendency of afterimages to fade in and out of conscious awareness (Wade, 1978), especially when not supported by a consistent context, could be analogous to the perceptual oscillations present during binocular rivalry. Increased ambiguity, whatever its source, is likely to result in greater susceptibility to contextual modulation. Moreover, in everyday viewing, such context would be powerful enough to suppress the perception of afterimages most of the time. Support for these suppression effects is found in the experiments of Daw (1962) and in the illusion by van Lier et al. (2009), where afterimage percepts are inhibited when the context was inconsistent with the afterimage (see also our “Welsh Castle” demo in supplemental information, where the afterimage is suppressed when the context is presented upside down). This might explain why we do not often perceive afterimages in everyday life despite the ease with which they can be evoked in demonstrations and illusions. 

Although we are not arguing that the underlying cellular populations differ between afterimages and perceptions of real chromatic objects, there are a number of plausible reasons (which we will discuss below) for why the pattern of activity in these cells may be importantly different between afterimages and real stimuli. If the perceptual system is attuned to these differences, the afterimage signals will present uncertainty—in some ways the signal will be like that of a real stimulus, but in some ways it will not be. In this situation, perception will be particularly influenced by any disambiguating cues, such as luminance edges, that increase the likelihood of the signal representing a real object. 

At a general level, our explanation relies on the assumption that it is beneficial for perception to dissociate afterimage signals from signals arising from real-world objects. Previous studies have shown that a colored surface is perceptually enhanced if it is interpreted as material (i.e., an object) rather than an illumination (Bloj et al., 1999; Kingdom, 2008). Afterimages could be considered similar to signal arising from illumination given that neither supplies direct information about objects. The visual system may have prior knowledge (implicitly, in the pattern and weights of its connections) of the typical activation profile associated with responses to real-world objects. If signals arising from adapted cells deviate in any way from the typical profile associated with real objects, this could lead to the interpretation that the afterimage signal is illusory. Thus, perception of the afterimage is suppressed. 

More specifically, we assume that luminance edges and other contextual cues support the interpretation that an ambiguous signal represents a real object. This interpretation may have been learned based on real-world statistics that most objects are delineated by luminance and chromatic contrast (Fine et al., 2003; Hansen & Gegenfurtner, 2009; Zhou & Mel, 2008), whereas this is not as likely for features less significant to awareness (e.g., illuminations, reflections). 

What is the source of afterimage ambiguity?

Although we have not explored directly ways in which the pattern of activation that produces an afterimage may be different from that arising from a real stimulus, it is possible to consider some likely candidates. One important factor that could disambiguate afterimages from real stimuli in normal viewing conditions is eye movements. Our eyes make frequent movements, and because afterimages are retinotopic they will move in exact synchrony with the eyes (Helmholtz, 1962). This characteristic is rarely (if at all) observed for real-world stimuli, even those followed with “smooth pursuit” eye movements (Kolarik, Margrain, & Freeman, 2010). Indeed, there is evidence that continual eye movements suppress the perception of afterimages (Kennard, Hartmann, Kraft, & Boshes, 1970; Matin, 1974). For these reasons, all our experiments were conducted with the eyes fixed to minimize the perceptual difference between afterimages and real stimuli created by eye movements. 

It is possible that even though we restricted larger eye movements in our experiments, small fixational jitter may have been sufficient to reveal that the afterimage percept was illusory. This knowledge could arise from compensatory mechanisms that stabilize the retinal image based on whole-world motion during jitter (Murakami & Cavanagh, 1998). These compensatory mechanisms would lead to a discrepancy between the afterimage (which is stabilized on the retina but has postcompensation movement) and the background (which moves on the retina but is stabilized postcompensation). 

The contour could have enhanced the perception of the afterimage because it provided a visual transient that is similar to the edge of a real object, and this transient would dominate the weaker signal from the afterimage itself. If jitter become too large, however, it is likely that it would dissociate the relative movement of afterimage and contour, and thus the enhancement effect would be lost. 

Apart from eye movements, we also controlled for many other possible perceptual differences between the afterimages and the physical stimuli (e.g., they were matched for hue, luminance, edge blur, and saturation). Controlling for these factors makes it possible to consider the existence of other critical factors differentiating the representation of afterimages from that of real stimuli of similar appearance. 

First, the temporal profile of an afterimage response is likely to be different from that representing a real stimulus. For example, the change from adaptor to afterimage (i.e., the off response) is unlikely to be exactly the same as the onset of a new stimulus. This difference could potentially underlie the experience of a small delay before the afterimage percept appears after the adapting stimulus is turned off (Creed, 1928). Similarly, the exponential recovery from adaptation (McLelland, Baker, Ahmed, & Bair, 2010) is unlikely to exactly mimic the activity profile displayed when viewing a real stimulus. Second, if adaptation is present at multiple color-sensitive sites throughout the visual system, cells in these sites may have different cellular architectures and different susceptibilities to adaptation and their relative recovery from adaptation could occur at different speeds (Fairchild & Reniff, 1995; Jameson, Hurvich, & Varner, 1979; Loomis, 1972; McLelland, Ahmed, & Bair, 2009; Rinner & Gegenfurtner, 2000; Yeh, Lee, & Kremers, 1996). This means that the relative firing rates after adaptation will differ between areas in a way that is inconsistent with how real stimulus signals elicit activity patterns through the visual system. 

How context could modulate chromatic signals

Perceptual demonstrations of filling in and psychophysical research both demonstrate that luminance edges are important for constraining and facilitating chromatic signals (Chaparro et al., 1994; Cole et al., 1990; Eskew et al., 1994; Gur & Akri, 1992; R. Hilz & Cavonius, 1970; R. L. Hilz et al., 1974; Montag, 1997; Mullen & Losada, 1994). In our experiment, a sharp edge was critical for enhancement above the mere presence of a luminance difference, as we observed that a blurry-edged pedestal did not enhance the afterimage as much as a sharp contour or pedestal. A luminance sine-wave grating does lower chromatic threshold, though not to the extent of square-wave gratings (Gowdy et al., 1999). That the edge closely frames the chromatic signal is also important, as orthogonal edges do not produce the facilitation observed with contiguous edges (Gowdy et al., 1999). Higher level edge representations, such as stereoscopic-depth edges and illusory contours, also modulate chromatic representations (Montag, 1997). This suggests that as long as the edge makes sense in terms of a higher level context it produces facilitation. It also seems important for facilitation that the edges are suprathreshold (Chaparro et al., 1994). 

The fact that chromatic facilitation is reliant on sharp edges, even if the properties of these edges are somewhat abstract, suggests that it is mainly a cortical process rather than just a low-level luminance interaction. This is in line with physiological research that orientation and form become increasingly important at higher levels of the visual system (Friedman et al., 2003) and that responses of blue–yellow color-opponent neurons in the macaque V1 are facilitated by luminance contrast (Horwitz et al., 2005). 

Loci of adaptation

The sites of adaptation that may contribute to generating color afterimages are worth discussing here as this has been the subject of much debate. Past research has variously championed the existence of afterimage-generating adaptation in the photoreceptors (Brindley, 1962; Craik, 1940), retinal ganglion cells (Virsu & Laurinen, 1977; Zaidi et al., 2012), and the cortex (Shevell, St Clair, & Hong, 2008; Shimojo, Kamitani, & Nishida, 2001). 

We can identify three possible sites of adaptation, although our results are consistent with any or all of these and do not exclusively point to a cortical locus. First, because we found that edges enhance afterimages more than they do physical stimuli, we could assume some adaptation at the level(s) at which edges enhance chromatic signals. As discussed earlier, this is likely to be cortical. Second, it is possible that although the edge first facilitates chromatic signals in the cortex, feedback signals are subsequently relayed back to the lateral geniculate nucleus (LGN) to amplify and tune its response (Ferster & Miller, 2000). This means that cortical influences that are responsive to the luminance edge, such as fine orientation tuning and attention, could evoke relative changes in LGN activity. Thus, adaptation responses in the LGN could be enhanced by edge-driven amplification by the cortex. Finally, higher level visual areas may be sensitive to temporal differences between retinal ganglion cell rebound signals resulting from adaptation and those arising from real stimuli. Thus, these ganglion signals may be more susceptible to edge enhancement because they are deemed more ambiguous or unusual by higher levels. As noted above, our results do not distinguish between these suggested adaptation sites, only that at some point the visual system is able to dissociate signals arising from adaptation from those representing real objects. 

Summary

We have presented converging evidence that luminance edges enhance afterimages more than they do physical stimuli of similar appearance. This finding appears to be specific to sharp edges, as a graded luminance pedestal did not produce the afterimage enhancement found with a sharp pedestal or contour. These results demonstrate that the brain processes signals arising from adapted cells nonequivalently to those arising from real stimuli. We suggest that because signals that generate afterimage percepts fail to perfectly match those arising from real stimuli, the visual system is unsure of whether the afterimage represents the presence of a real object. This would explain why afterimages are influenced by contextual cues that reduce their uncertainty more than responses triggered by real objects. Conversely, in everyday viewing contextual cues will be unlikely to align with afterimages, and thus our perception of them is often suppressed. 

Supplementary Materials

Docx S1 

Docx S2 

Demo 1 

Demo 2 

Demo 3 

Demo 4 

Demo 5 

Acknowledgments

Thanks to Tom Freeman and Chris Miles for useful comments on the manuscript. This work was funded by the BBSRC and Cardiff School of Psychology. 

Commercial relationships: none. 

Corresponding author: Georgie Powell 

Email: [e-mail] 

Address: School of Psychology, Cardiff University, Cardiff, UK.