Abstract: Accurate pain assessment can provide valuable guidance for patient analgesia. To address the problems of low efficiency, time-consuming, and laborious nature of traditional artificial pain assessment, a facial pain assessment algorithm is proposed based on VIT and CNN attention mechanisms to predict pain at multiple levels. Firstly, a facial pain expression acquisition platform is built to extract the video frame sequence and preprocess the data, establishing a pain expression database. Secondly, an improved multi-scale channel attention module is proposed to focus on key feature information. CNN and VIT are connected in parallel as the backbone network to extract more advanced facial local-global features, which are then fed to the long short-term memory network (LSTM) in a temporal manner for pain assessment. Finally, the model performance is verified on the pain expression database. The experimental results show that the precision, recall, F1-score and accuracy of the algorithm are 96.8%, 96.7%, 0.97 and 96.8%, respectively. Compared with other deep learning models, it can more effectively identify pain and make a certain contribution to pain assessment research in the field of rehabilitation.

Key words: pain assessment, facial expression, VIT network, convolutional neural network, attention mechanism

摘要： 准确的疼痛评估可以为病人镇痛提供一定指导，为解决传统人工疼痛评估效率低、费时费力等问题，提出一种基于VIT与CNN注意力机制的面部疼痛评估算法，对疼痛进行多级预测。搭建面部疼痛表情采集平台，提取视频帧序列并进行数据预处理，建立疼痛表情数据库；提出一种改进的多尺度通道注意力模块关注关键特征信息，将CNN和VIT作为主干网络并行连接，提取更高级的面部局部-全局特征，以时序方式输入长短期记忆网络（LSTM）进行疼痛评估；在疼痛表情数据库上进行模型性能验证，实验结果表明，该算法在精确率、召回率、F1分数、准确率指标方面分别达到96.8%、96.7%、0.97、96.8%，与其他深度学习模型相比可更有效识别疼痛，为康复领域疼痛评估研究做出一定贡献。

关键词: 疼痛评估, 面部表情, VIT网络, 卷积神经网络, 注意力机制