Deep learning enhanced attributes conditional random forest for robust facial expression recognition

Automated Facial Expression Recognition (FER) has remained challenging because of the high inter-subject (e.g. the variations of age, gender and ethnic backgrounds) and intra-subject variations (e.g. the variations of low image resolution, occlusion and illumination). To reduce the variations of age, gender and ethnic backgrounds, we have introduced a conditional random forest architecture. Moreover, a deep multi-instance learning model has been proposed for reducing the variations of low image resolution, occlusion and illumination. Unlike most existing models are trained with facial expression labels only, other attributes related to facial expressions such as age and gender are also considered in our proposed model. A large number of experiments were conducted on the public CK+, ExpW, RAF-DB and AffectNet datasets, and the recognition rates reached 99% and 69.72% on the normalized CK+ face database and the challenging natural scene database respectively. The experimental results shows that our proposed method outperforms the state-of-the-art methods and it is robust to occlusion, noise and resolution variation in the wild.

Automated summary

The study tackles the challenges of automated facial expression recognition by introducing a conditional random forest architecture and a deep multi-instance learning model. Experimental results demonstrate that the proposed method performs well on public datasets and exhibits good robustness in complex environments.