Related Experiment Video
Updated: Jan 17, 2026
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
5.3K
Robust emotion recognition for complex environments: ChildEmoNet model based on DETR-ResNet50 cascaded architecture
Zhang Shanshan1, Sha Yanlin2, Loy Chee Luen3
1Department of Early Childhood Education, Faculty of National Child Development Research Centre, Universiti Pendidikan Sultan Idris, Perak, Malaysia.
Plos One
|September 18, 2025
Summary
This study introduces ChildEmoNet, a novel framework for robust emotion recognition, even with facial occlusions. It effectively detects multiple people and extracts features, outperforming traditional methods in complex settings.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Affective Computing
Background:
- Facial occlusion poses significant challenges for emotion recognition systems in real-world environments.
- Traditional deep learning models often struggle with performance degradation under occluded facial conditions.
Purpose of the Study:
- To develop a novel cascaded emotion recognition framework, ChildEmoNet, for enhanced performance in complex environments.
- To address challenges in multi-person detection and discriminative feature extraction, especially under facial occlusion.
Main Methods:
- Proposed a cascaded architecture integrating Detection Transformer (DETR) for multi-person detection and ResNet50 for feature extraction.
- Developed specific robustness mechanisms for facial occlusion scenarios.
- Evaluated performance on categorical and dimensional emotion recognition tasks using the OMG Emotion Dataset.
Main Results:
- Achieved an AUC of 0.93 for standard emotion classification.
- Maintained 79% recognition accuracy under 30% facial occlusion.
- Obtained concordance correlation coefficients (CCC) of 0.52 (valence) and 0.46 (arousal).
- Demonstrated superior performance across varying lighting, face orientations, and partial occlusions.
Conclusions:
- The cascaded DETR-ResNet50 architecture effectively handles multi-person detection and feature extraction for robust emotion recognition.
- ChildEmoNet shows remarkable robustness in complex real-world conditions compared to traditional methods.
- This research advances emotion computing by providing a reliable solution for challenging environments.
