A ROI-guided deep architecture for robust facial expressions recognition

Abstract

This work proposes a robust facial expression recognition framework, focusing on discovering the region of interest (ROI) to train an effective face-specific of convolutional neural networks (CNN). By exploiting the relationships among ROI areas, the proposed deep architecture can improve the reliability of predicted targets. Our designed deep model is fine-tuned based on a pre-specified deep CNN instead of a new one trained from scratch. To increase the face expressions toward a robust deep CNN training, a novel data augmentation strategy called artificial face is designed. The performance of our deep architecture is evaluated on state-of-the-art databases such as CK+. To demonstrate the high generalizability of our approach, cross-database validations are conducted on the JAFFE and our own compiled Wild database. Comprehensive experiments have demonstrated the superiority of the method, i.e., achieving a recognition accuracy of 94.67% on the CK+ database, 53.77% on the JAFFE cross-database, 40.13% on the FER-2013 cross-database, and 37.25% on the Wild cross-database respectively.