This study presents a robust and adaptable framework for Facial Emotion Recognition (FER) by integrating optimized dataset selection with a structured pre-processing pipeline and a hybrid CNN-Transformer architecture. The research critically evaluates four widely used FER datasets CK-.v7, FER-2013, Face Expression Recognition Dataset, and JAFFE highlighting trade-offs in resolution, annotation quality, and demographic diversity. A comprehensive pre-processing pipeline, including grayscale conversion, histogram equalization, facial alignment, and occlusion handling, is implemented to enhance image clarity and feature consistency. Data augmentation techniques, such as random rotation and brightness variation, further simulate real-world variability. The proposed FER model combines Convolutional Neural Networks for local feature extraction with Transformers for capturing long-range dependencies, integrated via an occlusion-aware attention fusion mechanism. Extensive cross-dataset experiments and ablation studies validate the framework’s robustness, with significant accuracy improvements resulting from key pre-processing components. Grad-CAM visualizations and class-wise metrics reveal high performance on distinct expressions (e.g., happy, angry) and highlight challenges in distinguishing subtle emotions like fear and neutral. Overall, the findings demonstrate that the proposed hybrid framework, when supported by carefully curated data and pre-processing strategies, significantly improves generalization and real-world applicability of FER systems.