The Draw-a-Person (DAP) test has been a widely used practical instrument in psychological and developmental assessments to measure children’s cognitive development via human-figure drawings. Unfortunately, its traditional scoring process relies on manual inspections conducted by professionals, which is highly subjective and difficult to scale. In order to resolve these problems, this paper presents a hybrid approach that leverages deep-learning-based visual recognition and rule-based structural reasoning for automated evaluation of children’s DAP drawings. Specifically, the model assesses drawings based on 40 features, including anatomical parts, appearance-derived attributes, and high-level structural-drawing relations. A multi-label CNN built upon the ResNet-50 model predicts the visibles, and rule-based geometrical reasoning is adopted to infer structures, including attachments, proportions, symmetries, and placements. These two aspects are combined into a single hybrid representation yielding interpretable feature scoring consistent with developmental-evaluation standards. The proposed framework performs very well across multiple feature analyses, achieving a Micro-F1 of 95.32% and Macro-F1 of 91.72% on the test dataset, and demonstrating robust multi-label classification ability even on rare features. It provides a promising method for evaluating Draw-a-Person drawings, while offering reliable capabilities for feature analysis and scoring with accurate anatomical feature detection and reasonable structural and higher-level feature detection despite the challenging diversity of children’s drawing styles. The enforced rule-based structural reasoning improves interpretability and objectivity. Our future work includes extending the framework to cover further detailed DAP features.
Alwadai et al. (Fri,) studied this question.