Abstract
This paper explores the automation of spatial analysis in children's religious art. Using a dataset of over 6,500 drawings from the "Drawings of Gods" project, researchers evaluate machine learning techniques to localize the color yellow—a cross-cultural symbol for divine light. The study demonstrates that a 5x5 grid-based supervised classification using k-Nearest Neighbors (k-NN) provides a robust framework for interpreting the irregular spatial logic of children's drawings, outperforming traditional computer vision methods designed for natural images.
Introduction
In the psychology of religion, children’s artistic expressions serve as a window into their conceptualization of the divine. A recurring hypothesis in this field suggests that children systematically use the color yellow as a proxy for "light," which is frequently associated with divinity across various cultures. However, analyzing these patterns at scale presents a significant logistical challenge. The "Drawings of Gods" project has amassed a dataset of over 6,500 drawings, making manual annotation of color placement—whether yellow appears in the center, around a figure, or in the periphery—an inefficient task. This research seeks to bridge the gap between religious psychology and computer vision (CV) by developing an automated pipeline for yellow localization.
The Computational Challenge of Artistic Datasets
Standard computer vision techniques are typically optimized for "natural images" (photographs), which rely on consistent lighting, textures, and geometric structures. Children’s drawings, by contrast, are inherently "noisy." The strokes are irregular, color intensity varies based on the pressure applied to a crayon or pencil, and spatial composition does not follow the rigid rules of professional art. Consequently, traditional CV models often struggle to interpret the symbolic and irregular spatial logic found in these works.
Methodological Evolution
The researchers tested three primary computational approaches to replicate the accuracy of human manual annotation:
- Gravity of Intensity: This method attempted to calculate the "center of gravity" for yellow by averaging color intensity across rows and columns. While human observers could identify peaks in these averages, the high variability in how children apply color made it impossible to establish a universal mathematical criterion for these peaks.
- Circular Sectors: Building on the gravity model, researchers defined two concentric circles around the center of gravity, hypothesizing that the center of yellow would coincide with the center of the main figure. This approach failed because children’s drawings lack the geometric precision required for such radial models.
- The 5x5 Grid (Supervised Classification): The most successful approach involved dividing a squared version of each drawing into a 5x5 grid (25 cells). The intensity of yellow was calculated for each cell, and these 25 values served as features for multi-label supervised classification.
Among the algorithms tested, the k-Nearest Neighbors (k-NN) method yielded the highest accuracy in predicting manual labels, such as "yellow in the center" or "yellow in the periphery."
Data and Ground Truth
The study utilizes the "Drawings of Gods" dataset, an interdisciplinary collection funded by the Swiss National Science Foundation. The dataset includes over 6,500 drawings collected from children across diverse cultural backgrounds. To validate the automated models, a subset of these drawings was manually annotated to provide a "ground truth," specifying the location of yellow and the presence of specific elements like a yellow sun.
Key Insights and Discussion
The "Center" Bias
The research leveraged a known psychological heuristic: children tend to place the "main figure" of their drawing in the center of the page (Golomb, 1987). This finding informed the grid-based classification, as the spatial distribution of color is often relative to this central focal point.
Feature Engineering vs. Complexity
Rather than employing deep neural networks—which might overfit on such a specific, non-natural dataset—the authors found that simple spatial grids were more robust. This suggests that for artistic and symbolic datasets, structured feature engineering can be more effective than high-model complexity.
Limitations and Future Work
The primary limitation remains the non-constancy of composition. Children do not always center their subjects, and the physical medium (faint pencil vs. thick paint) affects intensity-based detection.
Future research will focus on:
- Multi-color analysis: Examining how yellow interacts with other colors to represent light or divinity.
- Adaptive Grids: Developing models that can identify the "main figure" first and then adjust the grid accordingly.
- Algorithm Refinement: Testing different parameters within supervised learning models to further decrease error rates across diverse artistic styles.
Conclusion
By transitioning from manual tagging to a grid-based k-NN classification, this project demonstrates a viable path for the large-scale analysis of children's art. This automated approach not only supports the hypothesis regarding yellow as a divine signifier but also provides a new methodological toolkit for researchers at the intersection of psychology and digital humanities.
Summary updated