Publication: Facial Action Units Recognition using Deep Learning Model with Bottleneck Features
Issued Date
2021-01-01
Resource Type
Other identifier(s)
2-s2.0-85125166549
Rights
Mahidol University
Rights Holder(s)
SCOPUS
Bibliographic Citation
ICSEC 2021 - 25th International Computer Science and Engineering Conference. (2021), 318-323
Suggested Citation
Konlakorn Wongpatikaseree, Narit Hnoohom, Sumeth Yuenyong, Sukrit Jaidee Facial Action Units Recognition using Deep Learning Model with Bottleneck Features. ICSEC 2021 - 25th International Computer Science and Engineering Conference. (2021), 318-323. doi:10.1109/ICSEC53205.2021.9684576 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/76714
Research Projects
Organizational Units
Authors
Journal Issue
Thesis
Title
Facial Action Units Recognition using Deep Learning Model with Bottleneck Features
Other Contributor(s)
Abstract
This paper presents a framework for detecting action units. The proposed framework consists of two frameworks: feature engineering for improving regression model accuracy and bottleneck features for improving deep learning model accuracy. For a regression framework, This paper presents an SVM with a feature formed by the feature engineering method, which consists of four features: histograms of directional gradients (HOG), facial landmarks, facial landmark distance from the center, and facial landmark angle relative to the center. For a deep learning framework, This paper presents AlexNet and VGG16 with bottleneck features and other optimization techniques such as hyperparameter tuning and image augmentation. Both frameworks presented were trained with the CK+ dataset. For the detection of 10 action units, the results showed that the deep learning framework outperformed the regression framework by 11%. For the detection of 17 action units, the results showed that AlexNet and VGG16 had the highest accuracy compared to other deep learning models, with an accuracy of 74%. Additionally, they outperform the linear SVC model and the Feed Forward MLP model by 10% and 12%, respectively.