Saurav, S and Kumar, T and Saini, R and Singh, S
(2020)
Video-Based Facial Expression Recognition using a Blend of 3D CNN and ConvLSTM.
In: 17th IEEE India Council International Conference (INDICON-2020), December 11-13, 2020, NSUT, New Delhi, India.
Abstract
The 3-Dimensional Convolutional Neural Network (3D CNN) and Long Short-Term Memory Network (LSTM) have consistently outperformed many approaches in video-based
Facial Expression Recognition (VEER). The vanilla version of the fully-connected LSTM (FC-LSTM) unrolls the image to a one-dimensional vector, which results in the loss of vital spatial information. Convolutional LSTM (ConvLSTM) overcomes this limitation by performing LSTM operations in terms of convolutions without performing any unrolling, as in the case with FC-LSTM. Motivated by this, in this paper, we propose a neural network architecture that consists of a blend of 3D CNN and
ConvLSTM. The proposed hybrid architecture captures spatial-temporal information to produce competitive accuracy on three publicly available FER databases, namely the CK+, SAVEE, and AFEW. The experimental results demonstrate excellent performance without using any external emotion data with an added advantage of having a simple model with a comparatively fewer number of parameters and model size. Our designed FER pipeline is a suitable candidate for automatic recognition of facial expressions in real-time on a resource-constrained embedded platform.
Actions (login required)
 |
View Item |
