Діагностика емоційного стану дітей на основі мультимодального інтелектуального аналізу малюнків

Abstract

This paper proposes a hybrid multimodal model for automated assessment of children’s emotional state from drawings. The approach combines EfficientNet-B3 global features, YOLOv8 symbol detection, HSV palette statistics and composition/graphical metrics. A lightweight attention-fusion layer weights channels and forms a 256-D embedding, which a two-layer MLP classifies into Happiness, Anxiety/Depression and Anger/Aggression. The attention weights quantify the contribution of each feature channel and support explainable decisionmaking. Experiments on a mixed test set (including Kaggle) reach 80–85% accuracy and produce an interpretable JSON report, which is useful for practical screening in educational and rehabilitation settings