Facial Expression Recognition through person-wise regeneration of expressions using Auxiliary Classifier Generative Adversarial Network (AC-GAN) based model

• Published in: Journal of visual communication and image representation Vol. 77; p. 103110
• Main Authors: V., Dharanya, Joseph Raj, Alex Noel, Gopi, Varun P.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.05.2021
• Subjects: Auxiliary Classifier GAN(ACGAN); Capsule Network(capsuleNet); Conditional GAN(CGAN); Facial Expression Recognition (FER); Subject dependence; U-Net; Conditional GAN(CGAN); Auxiliary Classifier GAN(ACGAN); Capsule Network(capsuleNet); Facial Expression Recognition (FER); Subject dependence; U-Net
• ISSN: 1047-3203; 1095-9076
• DOI: 10.1016/j.jvcir.2021.103110

Recently, Facial Expression Recognition (FER) has gained much attention in the research area for its various applications. In the facial expression recognition task, subject-dependent issue is predominant when a small-scale database is used for training the system. The proposed Auxiliary Classifier Generative Adversarial Network (AC-GAN) based model regenerates ten expressions (angry, contempt, disgust, embarrassment, fear, joy, neutral, pride, sad, surprise) from input face image and recognizes its expression. To alleviate the subject dependence issue, we train the model person-wise and generate all the above expressions for a person and allow the discriminator to classify the expressions. The generator of our model uses U-Net Architecture, and the discriminator uses Capsule Networks for improved feature extraction. The model has been evaluated on the ADFES-BIV dataset yielding an overall classification accuracy of 93.4%. We also compared our model with the existing methods by evaluating our model on commonly used datasets like CK+, KDEF. •Proposed a generative based model for facial expression recognition eliminating subject dependence issue.•Employed Capsule Network in the discriminator and U-Net in the generator of our model.•Our model recognizes four more expressions apart from the six basic expressions.•Used ADFES-BIV, CK+, and KDEF datasets to evaluate the performance of our model.•Various performance metrics and image quality metrics are provided for comparisons.