A facial action unit detection algorithm combining expert prior knowledge and motion correlation

Journal Title: Journal of Air Force Medical University - Year 2023, Vol 44, Issue 10

Abstract

Objective Facial action unit (AU) provides support for tasks such as expression recognition, micro-expression analysis, stress detection, fatigue detection, and lie detection. The purpose of this paper is to develop a stable and accurate facial AU detection algorithm based on integrating expert prior knowledge of face regions and AU relationships. Methods First, the key points of faces were used to divide each region of face images according to the expert prior knowledge of face regions, and the high-dimensional features of face images in each region were extracted with convolutional neural network. Then, the encoder in Transformer model was used to model the AU relationships and encode the high-dimensional features of faces in each region. Finally, an AU classifier was constructed using the fully connected layer. To tackle the problems of category imbalance and positive-negative sample imbalance in the dataset, a weighted differentiable F1 loss was incorporated into the loss function to constrain the AU detection model. Results In this paper, 3-fold cross-validation experiments were conducted on BP4D and DISFA datasets, and their average F1 scores for AU detection reached 63. 8% and 61. 8% , respectively. Conclusion The experimental results show that compared with the existing mainstream AU detection algorithms, the proposed algorithm has higher stability and detection accuracy on different datasets. The algorithm can accurately detect the facial AU, thus providing effective support for other tasks such as facial expression analysis.

Authors and Affiliations

LI Kui, MO Jianhua, WANG Jiajun

Keywords

Related Articles

Investigation and analysis of Chinese women􀆳s cognition of regular breast examinations

Objective To understand women􀆳s cognition and willingness for regular breast examinations through online surveys and analyze related influencing factors, so as to provide a reference for carrying out regular breast exa...

Effect of Gushen􀆳antai pill on endometrium and clinical outcome during frozen-thawed embryo transfer cycle

Objective To observe the effect of Gushen'antai pill on endometrium and clinical outcome during frozen-thawed embryo transfer cycle, and to explore its clinical application. Methods A total of 174 patients who underwent...

High altitude anti-hypoxia technology and equipment: research strategy based on man-machineenvironment system

Man-machine-environment system engineering is an interdiscipline in which human science and engineering technology penetrate each other. It ensures the optimal overall performance of man-machine-environment system by re...

Prevention and treatment of oxidative stress injury of vitiligo based on the TCM theory of “preventive treatment of disease”

The TCM theory of “preventive treatment of disease” can play a role in the prevention and treatment of many diseases. We found that the three states in the TCM theory of “preventive treatment of disease”, namely healthy...

Analysis and reflection on survey results for oral diseases in 420 officers and soldiers of a western unit

Objective To investigate the oral health status of grass-roots troops, and combined with the investigation results, put forward the diagnosis and treatment range of oral sergeants in grass-roots troops for reference. Me...

Download PDF file
  • EP ID EP723840
  • DOI -
  • Views 38
  • Downloads 0

How To Cite

LI Kui, MO Jianhua, WANG Jiajun (2023). A facial action unit detection algorithm combining expert prior knowledge and motion correlation. Journal of Air Force Medical University, 44(10), -. https://europub.co.uk/articles/-A-723840