The Efficient-CapsNet model for facial expression recognition

Facial expression recognition (FER) has attracted much attention lately. However, the current methods are concerned primarily with recognition accuracy, while ignoring efficiency. Efficient-CapsNet, which employs deep separable convolution operations based on CapsNet, has low network parameters and high network training efficiency while ensuring recognition accuracy. Using three public datasets, JAFFE, CK+, and FER2013, we comprehensively compared the recognition accuracy and training efficiency of Efficient-CapsNet and CapsNet. Results showed that the Efficient-CapsNet’s recognition accuracy reached 99.13%, 93.07%, and 72.94%, respectively, which is superior to most of the latest methods. In terms of training efficiency, the training time of a single image of Efficient-CapsNet under 64x64 size input and 48x48 size input is only 0.125ms and 0.033ms, respectively, which is 1454.28 times and 2730.03 times faster than CapsNet, respectively. Results also suggest that the training efficiency of Efficient-CapsNet is affected by the sample size. When the sample size grows, the training efficiency gradually slows down until it stabilizes.