Facial expression analysis is essential for monitoring academic engagement in dynamic classroom settings. However, existing models often struggle in real-time scenarios due to occlusion, pose variation, and the presence of multiple faces. To address these limitations, this study introduces UGen-RADis (U-shaped Generator and Residual Attention Discriminator Network), a novel GAN-based framework designed to enhance spatial feature extraction and ensure temporal consistency in facial expression recognition. The architecture incorporates a Full Skip-connection U-shaped Generator augmented by Frame Global Low-Frequency and Difference Local High-Frequency Attention mechanisms, while a Residual Attention Temporal Discriminator maintains coherence across video frames. Additionally, a 1D Logarithmic Gabor filter is employed for effective occlusion handling. The model was validated on real-world classroom videos involving 10–12 students and evaluated across six engagement categories: Boredom, Confusion, Frustration, Drowsiness, Neutral, and Engaged. Results show that UGen-RADis outperforms baseline models (LSGAN and CGAN), achieving up to 0.99 accuracy, 0.98 recall, and a 29% reduction in error rate. These outcomes highlight the potential of UGen-RADis for robust, scalable, real-time engagement monitoring in educational environments.