Facial expression recognition (FER) in real-world scenarios often faces challenges due to occlusions, such as masks, hands, or other objects, which significantly degrade the performance of traditional models. To address this issue, we propose an occlusion-aware FER framework that dynamically adjusts the importance of local facial features based on their relevance to the overall expression. Our approach introduces two novel modules: the Occlusion-Aware Module (OAM) and the Block Loss Module (BLM). The OAM leverages the cosine similarity between local and global features to dynamically weight the contribution of each local region, while the BLM optimizes the model by emphasizing the loss from non-occluded regions during training. In order to validate our approach, we construct a new dataset with varying levels of occlusion, derived from the RAF-DB dataset, and conduct extensive experiments on both occlusion and pose-variant datasets. Our results demonstrate that the proposed method outperforms existing state-of-the-art models, particularly in scenarios with moderate to heavy occlusions.