Facial Expression Recognition (FER) is fundamental to affective computing and human-computer interaction, enabling systems to infer emotional states from facial cues. Current deep learning approaches rely solely on raw facial images and standard augmentation but lack explicit incorporation of facial anatomical knowledge. This forces models to discover expressiondiscriminative regions from scratch rather than leveraging established facial behavior research. To address this gap, we propose MV-FER, a multi-view framework that integrates spatial priors from the Facial Action Coding System (FACS). MV-FER employs three data representations: original facial images, horizontally flipped images with random erasing for view consistency, and AU-guided heatmaps generated by mapping landmarks to established Action Unit areas providing explicit spatial guidance toward expression-relevant facial regions. The three views are unified within a training paradigm that encourages learning of discriminative features while maintaining robustness against pose variations and partial occlusions. Benchmark experiments demonstrate that MV-FER achieves enhanced recognition accuracy over state-of-the-art approaches with similar architectures. These results confirm the effectiveness of our multi-view strategy, with progressive accuracy gains as augmented and AU-guided heatmap views are incorporated.