Aphantasia is a condition in which individuals report either an inability to engage in mental imagery or report mental imagery that is lacking in vividness and clarity [28]. Although the condition may in some instances be multi-sensory (a subset of individuals with aphantasia report a lack of imagery across sensory modalities [8]), it has mostly been examined in the context of just one or a small subset of senses [5] with visual imagery being especially prominent [18]. Aphantasia can be assessed using subjective reports like the Vividness of Visual Imagery Questionnaire [21, 22], but recent research has focused on differences between individuals with aphantasia and those without on a variety of cognitive tasks. The goal of the majority of these studies is to establish objective perceptual and/or cognitive indices of the condition. For example, binocular rivalry is not affected by visual imagery instructions that do serve as a visual prime for control participants [18]. The absence of this priming effect suggests that individuals with aphantasia do not necessarily maintain a sufficiently vivid internal image of the prime stimulus to affect subsequent performance. In mental rotation tasks, individuals with severe aphantasia do not show the same effect of rotation distance on response time that control participants do, an effect that is typically interpreted as a reflection of some form of mental imagery used to manipulate an internal image of the target object(s) [23]. Despite this, individuals with aphantasia tend to perform as well as control participants in mental rotation tasks and other putatively visuospatial or working memory tasks that are assumed to depend on imagery, suggesting the adoption of some alternate strategy that does not depend on a pictorial representation of a target object that is absent [15, 20, 27]. In terms of physiological correlates of visual perception and visual imagery, individuals with self-reported aphantasia or reduced visual imagery as assessed by self-report also differ from controls in a manner consistent with objectively reduced or absent visual imagery. Kay et al. [17], for example, report that being asked to imagine achromatic patterns that are either bright or dark induces pupillary responses consistent with actually seeing these images if participants report vivid visual imagery, but this is reduced in participants reporting reduced vividness. During mental imagery, Zeman et al. also report that an aphantasic participant exhibited increased activity in frontal regions relative to control participants, coupled with reduced activity in parietal regions [27]. The parietal lobe contributes substantially to visuospatial processing, so reduced activity in this region is consistent with the absence of visual imagery and/or transformations of an internal image. Cortical excitability in areas V1–V3 is also lower as a function of imagery strength, as indexed using the binocular rivalry imagery paradigm in which instructions to imagine one member of the rivalrous stimulus pair affect subsequent rates of rivalry [19]. More broadly, imagery vividness in general correlates with the overlap between perceptual activity and neural activity during imagery, translating to a specific neural basis for imagery vividness in individuals with aphantasia and the general population [9]. Together, these results indicate that aphantasia does indeed reflect differences in visuospatial imagery between control participants and individuals subjectively reporting severely reduced visual imagery, both in terms of behavioral performance and neural responses.

We used the MLV toolbox to measure the frequency of different junction types in each drawing. In general, a junction refers to a feature where line segments meet or cross and these are classified by the number of segments that contribute to the junction and the relative orientation of the segments. The MLV toolbox supports the extraction of T, Y, X, Arrow, and Star junctions, all of which we opted to measure in the database images. In Figure 3, we display an example drawing from the database alongside the MLV coding of junction types across this image. We measured junction frequency collapses across orientation, yielding five bins per image (one per junction type).

For each set of feature distributions, we carried out our classification analysis using the k-nearest neighbors algorithm implemented in the JASP v3.0 Machine Learning module [16]. Each of the drawings made by all participants was included in the full data set for both our perception and recall analyses. We used a 50/50 split of the data to randomly assign data points to training versus test samples and held out 20% of the training data for validation. The number of nearest neighbors used for classification was optimized per analysis but capped at a maximum value of 10 neighbors.

With regard to the drawings made in the perception condition, we found that Junction Types, Contour Curvature, and Line Segment Orientation did not support robust classification according to participant group. The distribution of Junction Type features supported a test set accuracy of 52.6%, Contour Curvature features yielded a test set accuracy of 54.5%, and Line Segment Orientation resulted in a test set orientation of 51.3%. In all three of these cases, comparison against the distribution of values obtained from our permutation analysis suggested that these accuracies did not differ significantly from what we might expect by chance (50th, 80th, and 64th percentiles, respectively). We did find, however, that the distribution of Line Segment Length did support reliable classification of drawings by group membership with a test accuracy of 78.8%. Compared to the distribution of values obtained from our classification analysis, this accuracy was well above our 97.5th percentile cut-off.

The data from the recall condition was very similar to the results reported previously for the perception condition. In both cases, Junction Types, Contour Curvature, and Line Segment Orientation did not support robust classification according to participant group. The distribution of Junction Type features yielded a test set accuracy of 51.9%, Contour Curvature features yielded a test set accuracy of 58.3%, and Line Segment Orientation features yielded a test set orientation of 56.4%. Again, none of these values reached our percentile threshold determined from our permutation analysis, though Contour Curvature did reach the 90th percentile and Line Segment Orientation reached the 95th. Though these results are suggestive of some systematic differences between our participant groups, we do not discuss them further as they did not reach our criterion for significance. By comparison, the distribution of Line Segment Length supported much more robust classification of aphantasic versus control drawings: We observed a test set accuracy of 89.1%, which was far above our 99th percentile cut-off based on permutation test analysis. In Figure 6, we display the average distribution of Line Segment Length values in both the perception and recall conditions for both aphantasia and control participants.

Although we find that both distributions of line segment length allow for successful categorization of drawings by participant group, we see in the classification accuracy results and the plots in Fig. 6 that recall condition separability is particularly good. Indeed, the distributions of values in the perception condition are extremely close between our two participant groups, with only some subtle differences making accurate classification more likely. By comparison, the recall condition distributions are more clearly different and indicate a bias for individuals with aphantasia to include more longer line segments than shorter lines in their drawings compared to control participants. In Table I, we summarize the classification results from all feature sets in the perception and recall tasks.

We chose to continue with an exploratory classification analysis of the separability of drawings from the perception and recall tasks on the basis of the different mid-level feature distributions described earlier. For this analysis, we combined features from individuals with aphantasia and control individuals to focus on differences in line drawings resulting from direct observation versus memory. We used the same k-nearest neighbors parameters described previously and conducted separate analyses for each set of features. Like we observed in the previous analysis, we found that classification accuracy was much better when Line Segment Length distributions were used than any of the other feature sets. The distribution of Junction Type features yielded a test set accuracy of 59.3%, Contour Curvature features yielded a test set accuracy of 64.4%, and Line Segment Orientation features led to a test set accuracy of 62.2%. By comparison, accuracy with Line Segment Length led to a test set accuracy of 84.0%. We display the average distribution of line segment length measurements in both the perception and recall tasks in Figure 7.

Similar to the difference between individuals with aphantasia and control participants, drawings made in the recall condition tend to differ from those in the perception condition due to a bias favoring longer line segments in recall drawings. A preliminary analysis indicated that this was the case for both individuals with aphantasia and control participants, but due to the smaller sample size involved in these within-group comparisons, we are potentially unable to measure differences in classification accuracy robustly in this case.

Our results demonstrate that there are differences in mid-level image features found in drawings of natural scenes that depend on observers’ ability to engage in mental imagery. We also found that there are similar differences that are evident when comparing images made while directly observing a target scene to those made when attempting to recall that scene from memory. A unifying perspective on these two outcomes is that both the presence of aphantasia and the requirement to draw a scene from memory lead to relatively impoverished internal images. What we think is particularly interesting is that the lower fidelity of the internal images maintained under these conditions is only expressed robustly through one class of mid-level statistics considered here (namely, Line Segment Length distributions) and not through the others. This means that the previously reported effect of aphantasia on high-level features like object detail [2] does not manifest in some uniform cost to mid-level or low-level features that are associated with that detail. To put it another way, the fact that individuals with aphantasia tend to make sparser drawings from memory than control participants does not impact all simpler visual features in the image. Instead, the vividness of visual imagery only affects the expression of a limited class of mid-level structures.

First, why should Line Segment Length statistics be impacted in the way that we have observed as a function of aphantasia and the imposition of a recall constraint? The bias that we observed in both conditions that we are associating with a weaker internal image favored longer line segments over shorter line segments, which could be the result of a number of pictorial differences. Shorter line segments tend to contribute to texture patterns, small-scale objects, and surface details, or could be used to render complex external outlines of objects or surfaces via piecewise drawing of an extended contour. One feature of the Aphantasia Drawing Database that is interesting to consider here is the use of text to label objects in scenes (e.g. putting the word “table” inside an otherwise indistinct shape), which was more prevalent in drawings by individuals with aphantasia. This practice should inflate short line segments relative to long ones, but obviously does not do so to a sufficient degree to overcome this more pronounced bias. Instead, the bias favoring longer contours in aphantasic drawings and drawings made from memory likely reflects the lack of surface detail in particular. Piecewise drawing of complex contours would likely result in orientation histogram differences that we do not observe in our data, leaving the presence (or relative absence) of surface detail as the main contributing factor. We note that this need not be directly tied to object detail, but may be a more general feature of how surfaces are rendered under these different conditions.

Besides asking why one of our feature classes is impacted by these conditions, we can also ask why the others are not. Though the other classes of mid-level features were better at separating perception drawings from recall drawings than at separating aphantasic versus control drawings, accuracy in the former classification analysis was still fairly low. One possibility that to our knowledge has not been explored is that some of the feature distributions we considered here may be lawful in the sense that constraints on the nature of line drawings lead to a narrow space of possible shapes for these distributions. By analogy, we are proposing something akin to the known properties of photographs of natural scenes like the approximate 1∕f power spectrum of natural images [12] or the distribution of cardinal orientations relative to oblique orientations in natural scenes [3]. The former property of natural scenes is a consequence of the scaling of 3D shapes and surfaces at different positions when projected into two dimensions [24], while the latter is a function of gravity favoring contours either parallel to the ground plane or perpendicular to it [4]. Any lawful properties of line drawings would not be subject to the same kind of physical and optical constraints but might instead reflect motor constraints or biases on drawing. A broader examination of how these mid-level statistics may vary across other types of drawings or how they may vary developmentally may yield important insights into how curvature, orientation, and junctions are typically expressed in line drawings of complex scenes. Though there have been several recent discussions of why line drawings “work” perceptually [14], what we are suggesting is a close analysis of how line drawings tend to be made by both trained and naïve participants. As Sayim and Cavanaugh observe [25], there have been surprisingly few changes in some aspects of line drawing production over very long periods of human history, which may indicate the presence of something like the natural modes in line drawing statistics we are suggesting as the basis for the null results in these different feature classes.

Our analyses do have some limitations that are also important to consider alongside our discussion of these interesting aspects of our results. One aspect of the database that we did not include as an additional factor is the different images created by each participant of the three different indoor scenes. The advantage of the aggregate analysis presented here is that we end up with thrice the number of data points, but the downside is that there is within-participant variability we are not considering. On one hand, there may be interesting individual variations in how some of these mid-level features are expressed, which could potentially be examined in a mixed-model analysis that includes image as a factor. To the extent that individual participants do draw with a unique style that is evident in the feature distributions we considered here, this could inflate our classification accuracy somewhat given that other drawings made by the same participant may be particularly useful nearest neighbors for classification. There are also parameter values regarding the normalization of these different features by contour length, the granularity of feature histograms, and other properties of mid-level feature measurement that we did not explore in depth here. Although we do not anticipate that our results would be sensitive to these parameters, it is potentially worth investigating the robustness of our results along these lines. Finally, our choice of classifier was motivated largely by a desire for simplicity rather than maximizing our classification accuracy. It is possible that discriminant methods, including SVMs, could yield higher classification rates for our other feature distributions. Again, we doubt that this is likely to be the case, but we have not explored this possibility in any depth. Nonetheless, despite these avenues for further refinement of our approach, we think that the current study offers new and interesting insights into how impoverished internal images do and do not lead to effects on mid-level visual statistics.

Attempting to reproduce natural scenes via line drawings with an internal image that is not especially vivid leads to differences in the mid-level statistics of line length favoring longer line segments. This result indicates that both aphantasia and drawing from memory lead to reduced fidelity of surface detail and other small-scale features in complex scenes. The absence of effects on other aspects of mid-level visual statistics may suggest either that these properties of natural scenes are robust to the strength of an internal image or that these features are constrained in some manner by the nature of line drawing as a perceptual task and/or a motor task. Examining mid-level characteristics of line drawings more broadly may reveal lawful properties of line drawings or other differences in the structure of line drawings made under different constraints and of different kinds of natural objects, surfaces, and scenes.