Robust emotion recognition for complex environments: ChildEmoNet model based on DETR-ResNet50 cascaded architecture

Abstract

Emotion recognition faces significant challenges in complex real-world environments, particularly under facial occlusion conditions that severely impact traditional deep learning approaches. This research proposes ChildEmoNet, a novel cascaded emotion recognition framework that strategically integrates Detection Transformer (DETR) for robust multi-person detection with ResNet50 for discriminative feature extraction. The primary contributions include the development of a cascaded DETR-ResNet50 architecture that addresses both detection and classification challenges simultaneously, enhanced robustness mechanisms specifically designed for facial occlusion scenarios, and comprehensive evaluation across both categorical and dimensional emotion recognition tasks. Extensive experiments on the OMG Emotion Dataset demonstrate the effectiveness of this integration: the proposed model achieves an AUC of 0.93 in standard emotion classification tasks, maintains 79% recognition accuracy under 30% facial occlusion conditions, and attains concordance correlation coefficients (CCC) of 0.52 and 0.46 for valence and arousal prediction, respectively. The experimental validation confirms the crucial role of the DETR module in processing multi-person scenarios and the effectiveness of ResNet50 in feature extraction, demonstrating superior performance across complex environmental conditions including varying lighting, face orientations, and partial occlusions. Compared with traditional methods, this cascaded architecture shows remarkable robustness under challenging real-world conditions. This research advances emotion computing technology by providing a robust solution for emotion recognition applications in complex environments where conventional approaches exhibit significant performance degradation.