Comparative Analysis of Hybrid Face Detectors and Classifiers for Child Emotion Recognition and Toy Recommendation

Abstract

The rapid advancement of affective computing has opened new avenues for supporting early childhood development through intelligent systems. However, traditional facial expression recognition (FER) models, primarily trained on adult datasets, often fail to generalize to the unique facial structures and nuanced emotional displays of infants and toddlers. This research presents an integrated AI-driven framework designed to provide personalized toy recommendations based on real-time infant emotion detection. The system utilizes a specialized pipeline incorporating MTCNN for precise facial alignment and a comparative analysis of high-performance architectures, including EfficientNet-B0, DenseNet121, ResNet50, and YOLO for robust feature extraction and classification. To address the critical scarcity of open-access data for the 0–1.5 year age group, a domain-specific dataset was curated to enhance model reliability in detecting subtle affective states. Beyond technical classification, the research introduces a scalable mobile application that bridges the gap between emotion analysis and consumer intelligence by offering mood-aligned toy suggestions with integrated price-tracking features. Experimental results indicate that our approach significantly provides a high degree of accuracy in resource-constrained environments. This study contributes a foundational methodology for infant-centric AI applications, offering a practical tool to foster emotional growth and personalized play experiences on a global scale.