Capsule Network-Based Facial Expression Recognition Method for a Humanoid Robot

摘要

Compared to the classical convolutional neural network (CNN), the capsule net Hinton put forward can use fewer network layers to achieve the classification tasks very well and arrive at the convergence with a faster speed. The principle of the capsule net is based on the CNN, and it is just that the neuron form is converted from the scalar to the vector, which is a capsule, and then chooses the suitable capsule for the final output through the dynamic routing method (Sabour in Dynamic routing between capsules, [1]). In this paper, on the basis of the capsule net, use deconvolution to restore images and optimize the error between original images and restored images. The classical facial emotions database named Cohn-Kanade Database Plus (CK+) that is processed through Data Augmentation is used to conduct experiments. Lately, the classification results are combined with the NAO robot. The NAO robot is able to visualize the emotion by changing its eyes colors and speaking the results, achieving the purpose of combining theory with practice.