To address the limitations of conventional facial emotion recognition (FER) models—namely inadequate precision in modeling expression-related features, insufficient responsiveness to subtle micro-expression regions, and constrained computational resources for mobile deployment—this paper proposes a lightweight yet high-performance FER model termed LMFER (Lightweight MobileViT for Face Emotion Recognition). Built upon the MobileViT backbone, LMFER performs structural optimization by introducing an improved Coordinate Attention (CA) module and an Emotion Attention Mechanism (EAM) into the first two network stages, respectively. The CA module enhances salient-region extraction from both spatial and semantic perspectives via directional spatial encoding and the incorporation of a channel-attention component (i.e., the Channel Attention in CBAM), enabling joint modeling across spatial and channel dimensions. The EAM strengthens the responses to fine-grained, emotion-critical regions (e.g., eyebrows, eyes, and mouth corners) by generating multi-channel emotion maps and applying a Softmax-based normalization strategy. Moreover, EAM supports the integration of facial landmarks as semantic priors to guide attention toward expression-discriminative areas more precisely. In addition, Dropout regularization is introduced after the fully connected layer to mitigate overfitting, and training stability as well as generalization are further improved by adopting the AdamW optimizer and a dynamic learning-rate scheduling strategy. Experimental results demonstrate that LMFER achieves an accuracy of 95.08% on the RAF-DB dataset, outperforming MobileViT by 4.78%, 4.55%, and 4.67% in precision, recall, and F1-score, respectively, while maintaining only 2.1M parameters. LMFER also exhibits clear advantages over mainstream approaches such as ResNet50, MobileNetV2, ShuffleNetV2, GoogLeNetV1, and VGG-16. Overall, LMFER delivers notable improvements in recognition performance while retaining favorable model compactness and deployment adaptability, indicating strong potential for practical applications in human–computer interaction and affective computing.lease download TEMPLATE HELP FILE from the website.