An Improved ShuffleNet with SE Attention and Focal Loss for Lightweight Facial Expression Recognition

Facial expression recognition (FER) plays a critical role in human-computer interaction, mental health monitoring, intelligent security inspection and classroom emotional computing. Traditional deep convolutional neural networks for FER suffer from excessive computational overhead and large parameter volume, which limits their deployment on mobile terminals and embedded devices with limited computing resources. As a classic lightweight network, ShuffleNet adopts group convolution and channel shuffle operations to reduce model complexity, yet it lacks the ability to capture subtle facial emotional features and cannot well handle the class imbalance problem existing in mainstream FER datasets. To address these defects, this paper proposes an improved lightweight ShuffleNet model for facial expression detection. First, Squeeze-and-Excitation (SE) channel attention modules are embedded after each shuffle block to dynamically recalibrate feature channel weights and strengthen the extraction of discriminative facial expression features. Second, a multi-dimensional data augmentation strategy combining geometric transformation, color jitter and Gaussian noise injection is designed to expand sample diversity and enhance the model’s generalization ability under complex lighting, shooting angles and background interference. Third, Focal Loss is introduced to replace the standard cross-entropy loss, which suppresses the loss contribution of easy-classified majority samples and forces the model to focus on hard-to-distinguish minority facial expressions such as anger and disgust. Comprehensive experiments are conducted on three public datasets FER2013, CK+ and AffectNet. The results demonstrate that the proposed model achieves higher recognition accuracy compared with original ShuffleNet V2 while maintaining low computational cost and small parameter size, and presents superior robustness against unbalanced data and complex real-world scenes.