A multi-view face detection and expression recognition method with improved RetinaFace
Rui Zhong, Bin Jiang, Nanxing Li, Qinggang Wu, Hua-Wen Chang · International Conference on Mechanisms and Robotics (ICMAR 2022) · 2022
To address the problems of large detection error and low recognition accuracy of multi-view face expression recognition, this paper proposes a multi-view face detection and expression recognition method based on RetinaFace. Firstly, it relies on ResNet-50 as the backbone feature extraction network of RetinaFace for multi-view face detection. This effectively prevents overfitting problem caused by the heavyweight network ResNet-152 and the false detection problem caused by the lightweight network MobileNet-0.25 in the original RetinaFace algorithm. Secondly, the multi-task loss function is adjusted to ensure accurate detection rate while speeding up the detection rate. Finally, the detected face images are fed into the ResNet-50 network for expression feature extraction and classification. In comparison with the baseline algorithm, the improved RetinaFace algorithm shows good robustness in terms of detection accuracy as well as detection time; it also shows good generalization in terms of expression recognition rate on the Multi-PIE facial expression datasets.