Abstract
Facial Expression Recognition (FER) has emerged as an important affective phenotyping tool in computational psychiatry, providing objective measurements of facial expressivity associated with psychomotor abnormalities in psychiatric disorders. From the perspective of smart product design, deploying deep learning frameworks for longitudinal psychiatric monitoring requires balancing recognition accuracy with hardware constraints. Since existing dual-stream networks often rely on computationally intensive global cross-attention, their integration into resourceconstrained products or interactive terminals remains challenging. To address these limitations, we present KGLA-Net, a dual-stream network. Our framework extracts parallel representations through a landmark backbone and an image backbone across three hierarchical stages. In the structural stream, a feature refinement module processes emotion-relevant cues, which are then aggregated through max-pooling to form multi-scale refined keypoint features. This structural prior drives a Keypoint-Guided Local Aggregation (KGLA) module, employing multi-head cross attention across localized multi-scale windows (8 × 8, 4 × 4, and 2 × 2) to guide the aggregation of image features, thereby reducing the computational complexity to linear complexity. Finally, the aggregated features are processed via keypoint feature self-attention before interacting with a zero-initialized query in stacked attention blocks. Experimental evaluations demonstrate that KGLA-Net achieves competitive FER performance while maintaining high computational efficiency. The proposed framework provides an efficient solution for facial affective computing and offers a lightweight foundation for future research on computational psychiatry.