Facial Paralysis Evaluation Based on Improved Residual Network

The evaluation of facial paralysis usually depends on doctors' clinical experience and on-site observation, which has subjective uncertainty and irreproducibility. In recent years, various computer-aided diagnosis techniques have been widely used in facial paralysis recognition and evaluation. Manual screening is required for facial feature selection. However, this may result in the inaccuracy in facial paralysis evaluation. To address this issue, a facial paralysis recognition method based on improved residual network model (multi-scale convolution based ResNet, MC-ResNet) by using 2D facial images is proposed in this paper. Multi-scale convolution and convolutional block attention module (CBAM) mechanism are introduced into the residual network to extract multiple scales and detail facial features. The spatial features and channel features are weighted to extract multi-level facial information. The experimental results trained by public Youtube Facial Palsy Database(YFP) show that the recognition accuracy of the improved residual network combined with the multi-scale convolution and CBAM mechanism is higher than that of the traditional deep learning model in the evaluation of facial paralysis grade.