Optimized Dense Convolutional Neural Networks for Micro-expression Recognition

Abstract

Micro-expressions are facial expressions that can reflect genuine human emotions. Alas, manual recognition of micro-expression is a time-consuming and arduous task due to its low-intensity reactions and brief occurrence. Convolutional layer, which is a well-known component in a deep learning architecture, are often used to learn the micro-level expression features so that the right micro-expression can be recognized. However, there is bound to be some feature loss when the feature maps are down-sampled towards the end of the network. If the loss occurs in the early layers, the network capability to learn the optimal features will be affected, which in turn degrades the model performance. In this paper, pooling layers are placed at the later layers, rather than the early layers to ensure optimal feature learning. In addition, a new set of hyperparameters are fine-tuned to deal with the learning problems caused by the modified pooling layers. For further improvement, the residual skip connections are also fed to forward layers, which are then combined using concatenate operator. The models require an input set of micro-expression onset-apex optical flow features to learn and recognize the correct emotion class; namely positive, negative, and surprise emotions. The overall recognition accuracy of micro-expression recognition has improved by around 4.83% compared to the base model. Hence, the proposed network improvements and modifications have managed to better recognize the correct micro-expression.