Mobile Convolutional Neural Networks for Facial Expression Recognition

We propose CNN models for facial expression recognition that work well in mobile and embedded devices. Previous studies introduced CNN models for image classification by stacking wider filters in depth to increase accuracy. The deep CNN models improve classification accuracy, but it is difficult to use in mobile devices because of its large parameter size and low responsiveness. We first analyzed the MobileNetV2 for facial expression recognition in mobile devices. After that, we designed CNN models with less than 1 million parameters by adjusting the width and depth of the bottlenecks. We trained the proposed CNN models and other mobile CNN models under the same experimental conditions and reviewed the results. The proposed CNN models have been carefully fine-tuned to use less than 0.5 million parameters. The fine-tuned CNN models achieved an accuracy of 90.3% for 5 classes and 86.8% for 7 classes in the RAF database.