A Fast Real-time Facial Expression Classifier Deep Learning-based for Human-robot Interaction

Human-robot interaction drives the need for vision technology to recognize user expressions. Convolutional Neural Networks (CNN) has been introduced as a robust facial feature extractor and can overcome classification task. However, it is not supported by efficient computation for real-time applications. The work proposes an efficient CNN architecture to recognize human facial expressions that consist of five stages containing a combination of lightweight convolution operations. It introduces the efficient contextual extractor with a partial transfer module to suppress computational compression. This technique is applied to the mid and high-level features by separating the channel-based input features into two parts. Then it applies sequential convolution to only one part and combines it with the previous separated part. A shuffle channel group is used to exchange the information extracted. The structure of the entire network generates less than a million parameters. The CK+ and KDEF datasets are used as training and test sets to evaluate the performance of the proposed architecture. As a result, the proposed classifier obtains an accuracy that is competitive with other methods. In addition, the efficiency of the classifier has strongly suitable for implementation to edge devices by achieving 43 FPS on a Jetson Nano.