Facial expression recognition (FER) holds significant application value in fields such as human-computer interaction and mental health analysis. However, persistent challenges remain due to subtle expression differences and complex real-world conditions. This paper proposes the Multi-Type Regional Attention Network (MTRAN), an innovative deep learning framework enhancing FER robustness. The proposed model architecture consists of four main components: 1) a backbone for global image feature extraction, 2) a Four-Region Feature Decomposition (FRFD) module to spatially segment the global feature map into four non-overlapping regions, 3) a Multi-Type Regional Attention (MTRA) mechanism to model intricate inter- and intra-regional dependencies, and 4) an Efficient Channel Attention (ECA) block to refine feature representation. To fully leverage the extracted global feature map, we spatially decompose this into four non-overlapping local regions. Our MTRA then captures both the inter-regional correlations among these local regions and their intra-regional dependencies. The ECA block adaptively emphasizes informative channels while suppressing redundant ones, further refining regional features. We conducted extensive evaluations on three public datasets, achieving state-of-the-art accuracies of 92.83% on RAF-DB, 67.69% on AffectNet-7, 64.38% on AffectNet-8, and 93.81% on CAER-S. Compared to existing models, MTRAN demonstrates superior performance. Code will be released at https://github.com/Sorenky/MTRAN