Research on Micro Expression Emotion Recognition Algorithm Based on Improved YOLOV8 and ResNet

With the rapid advancement of artificial intelligence, micro-expression recognition has demonstrated significant application value in fields such as criminal investigation and lie detection. However, the extremely short duration of micro-expressions and the scarcity of annotated data pose substantial challenges to accurate recognition. This study proposes a two-stage micro-expression recognition method based on an improved YOLOv8 and ResNet framework. In the first stage, a YOLOv8 model enhanced with StarNet multi-scale feature fusion and the CBAM attention mechanism is employed to achieve high-precision face detection. In the second stage, the detected static facial frames are fed into a ResNet50 network embedded with an SE module to strengthen channel-wise feature representation, thereby enabling micro-expression classification. For experimental validation, the improved models were trained and tested on the WiderFace and RAF-DB datasets. Results indicate that the improved YOLOv8 achieved an mAP of 0.693 on WiderFace, representing an increase of 0.109 over the baseline, while the SE-ResNet50 attained an F1-score of 0.7548 on RAF-DB, with an improvement of 0.2. These findings confirm the effectiveness of the proposed approach in both face detection and micro-expression classification. Moreover, this study provides a feasible solution to address the challenges of capturing micro-expressions and data insufficiency, while also laying the groundwork for future research on end-to-end joint modeling and lightweight network optimization.