Feature Pruned and Weighted Convolutional Neural Network for Micro-Expression Recognition

Micro-expressions are brief, involuntary facial movements that reveal concealed emotions. While Convolutional Neural Networks (CNNs) are effective for recognizing such expressions, they are often computationally intensive and memory heavy. To address this, a Feature-Pruned and Weighted Convolutional Neural Network (FPW-CNN) is proposed to compress and accelerate performance. This model prunes filters and selects only the most relevant responses, reducing memory usage and computational load. Instead of stacking outputs, convolutional layers are flattened. Experiments on five spontaneous micro-expression datasets—SMIC, CASME, CASME II, CAS(ME)2, and SAMM—show that FPW-CNN achieves accuracies of 99.75%, 99.93%, 99.95%, 99.94%, and 99.90%, respectively. The model also reduces FLOPs by over 41% and maintains over 95% accuracy at 128×128 resolution, demonstrating robustness. These results confirm that FPW-CNN outperforms recent CNN and transformer models, making it suitable for real-time micro-expression recognition on resource-constrained devices.