Emotion Analysis of Children’s Drawings

This study investigates children’s emotional expression through the computational analysis of their drawings. Prior research has frequently relied on proprietary or small-scale datasets, often limited to digital drawings produced on tablets or smartphones, which may not fully capture the natural characteristics of hand-drawn expression. To address these gaps, we introduce KIDO, a novel publicly available dataset of scanned children’s drawings paired with reflections written by the children on whether their drawings represent happiness or sadness. The dataset contains contributions from 5,430 students, with balanced representation across primary and secondary education levels and additional metadata on gender and educational background. For our experiments, we employed CNN-based and Vision Transformer–based models for emotion recognition, obtaining F1-scores of up to 79.03% from drawings, while BERT-based models achieved accuracies of up to 97.47% in classifying children’s written reflections. We further trained a generative model on the KIDO dataset to synthesize emotionally expressive drawings, with subjective evaluations indicating that children often preferred the AI-generated images to those drawn by their peers. KIDO serves as a comprehensive benchmark for advancing research in the fields of computer vision, natural language processing, psychology, and education, and is publicly available at github.com/serdarciftci/KIDO