3D-Based Facial Emotion Recognition using Depthwise Separable Convolution

Facial images play a significant role in expression prediction. The 3D features of facial expression provides significant information. In the area of facial emotion recognition, the 3D geometry and 2D texture helps to improve the recognition rate. A lot of research works had achieved state-of-the-art results using handcrafted and deep convolutional neural networks containing many trainable parameters which require high computing power. In this paper, we employ two kinds of convolutions i.e., regular or normal convolution on the 2D texture image and separable convolution on the 3D depth map images. We run experiments with our proposed network on the BU-3DFER database. The proposed model was trained from scratch to adjust the weights and biases of the learnable layers on various image features and achieved state-of-the-art accuracy of 81.81% on the 2D texture image, 79.10% recognition accuracy on the 3D depth map, and 83.01% for combined 2D and 3D features.