Real-Time Non-invasive Mental Health Monitoring System Integrating Micro-expression Recognition and Localized Quantized Large Model

Aiming at the pain points in traditional mental health monitoring such as stigma hindering help-seeking, large errors in subjective reports, delayed intervention response, and privacy leakage risks, this study combines affective computing and computer vision technologies to develop a real-time non-invasive auxiliary mental state detection system based on the edge-cloud collaboration architecture. The system adopts a full-link design of “image acquisition - preprocessing - face detection and alignment - emotion recognition - localized large model feedback”, with the public FER2013 dataset as the core verification benchmark. Data augmentation strategies such as horizontal flipping and brightness adjustment are used to solve the sample imbalance problem, and it is synchronously adapted to intelligent acquisition devices with frame rate ≥30 fps and resolution ≥720p. In the preprocessing stage, MTCNN face detection and Dlib 68-point localization are used to crop ROI, and normalization is combined to adapt to the dataset format, reducing computational complexity by 80%. The core of emotion recognition is the lightweight LightEmotionNet model, supplemented by the EmotionCNN enhanced model to cover multiple deployment scenarios. The 4-bit quantized llama3:8b large model is deployed locally through Ollama, generating personalized “empathy + suggestion” feedback in 1.6 ± 0.3 s to eliminate data leakage. The system’s end-to-end latency is 1187 ms, providing reliable technical support for mental health monitoring and early warning.