Facial Expression Recognition (FER) is a key technology for understanding human emotions in applications such as human–computer interaction, healthcare, and safety monitoring. However, most existing studies assume ideal conditions with fully visible and frontal faces, which are often violated in real-world scenarios. In practice, facial images are frequently affected by occlusion (e.g., masks, sunglasses) and pose variations, which significantly degrade recognition performance.

This research aims to systematically analyze the effects of occlusion and pose variation on FER and to develop robust and lightweight recognition methods under such conditions. The problem is decomposed into three controlled settings: occlusion only, pose variation only, and their combination, enabling structured analysis of performance degradation.

The study utilizes both static image datasets and video-based datasets, with controlled occlusion modeling and pose-based grouping. Lightweight convolutional neural networks are adopted as baseline models, and different strategies such as local region importance estimation and pose-related auxiliary information are investigated.

Evaluation is conducted using standard metrics including Accuracy, Macro F1-score, and Weighted F1-score, with particular attention to class imbalance and robustness. The expected outcome is to provide systematic design guidelines for robust and efficient FER systems, extending from static image analysis to video-based and real-time applications.