Asosiy kontentga oʻtish
proceeding

Facial Emotion Recognition Enhancement Using CNN-RNN Hybrid Model in Physical Education Analytics

Najmitdinov Akhadkhon KhamitdkhanovichTuran International UniversitySaboxat KabulovaMamun UniversityС RajanK S Rangasamy College of TechnologySwetha S VSt. Joseph's Institute of TechnologyR UdayakumarKalinga University
2026en
Ochiq kirishOltin

Litsenziya: cc-by

ABI

Annotatsiya

Facial emotion recognition (FER) in physical education (PE) contexts presents a technically demanding variant of the general FER problem: PE environments combine continuous motion blur, lateral and downward face orientations from physical activity, variable lighting in indoor and outdoor facilities, and rapidly transitioning emotional states driven by effort, achievement, and social interaction that no single-frame convolutional model adequately captures. This paper presents a CNN-RNN hybrid model for PE-context FER that combines a ResNet-50 convolutional backbone for spatial feature extraction from individual frames with a Bidirectional Long Short-Term Memory (BiLSTM) recurrent network for temporal sequence modelling across consecutive frames, addressing the motion and transition characteristics of PE affective states. The model is trained and evaluated on a purpose-built PE-context facial emotion dataset of 48,400 video sequences (8 consecutive frames per sequence; 387,200 total frames) from 2,840 physical education students across 18 secondary schools in three countries (United Kingdom: 8 schools; Australia: 6 schools; Singapore: 4 schools), annotated for seven emotion classes (Neutral, Happy, Sad, Angry, Fearful, Surprised, Disgusted) by trained psychologists achieving inter-rater reliability$\kappa=0.84$. The CNN-RNN hybrid achieves an overall accuracy of 88.4%, weighted$\mathbf{F 1}=\mathbf{0. 8 7 2}$, and$\mathbf{A U C}-\mathbf{R O C}=\mathbf{0. 9 4 2}$on the stratified 20% held-out test set ($\mathbf{n}=\mathbf{9, 6 8 0}$sequences), outperforming ResNet-50 single-frame (78.4%), VGG-16 (76.4%), LSTM-only (72.4%), and the FER2013 public benchmark baseline (68.4%). Per-class analysis reveals that Happy achieves the highest recognition accuracy (94.8%) and Disgusted the lowest (78.4%), consistent with the PE context in which disgust is the least frequently exhibited and most contextually ambiguous of the seven target emotions. Applied as a real-time PE analytics system across 18 schools, the model achieves 84.6% accuracy in detecting student engagement states and 82.4% accuracy in fatigue detection, enabling instructors to adapt session pacing and task difficulty based on moment-to-moment affective feedback.

Hali tarjima qilinmagan

Identifikatorlar

Iqtiboslar va manbalar