Facial Expression Recognition: A Comparison of Bottleneck Feature Extraction

Abstract

This paper compares the performance of bottleneck feature extraction based on two different architectures, the first, Convolutional Neural Network (CNN) based bottleneck feature extraction and the second, Deep Neural Network (DNN) based bottleneck feature extraction. Both of CNN and DNN based bottleneck feature extraction network were trained for 200 epochs to perform a feature extraction task. The input of bottleneck network is the same as the output which is the preprocessed images. From the bottleneck network, after training, the layers after the bottleneck layer were cut-off and set the bottleneck layer as an output layer. The result of the bottleneck feature extraction is that it can reduce the dimension of the images from 4096 to 128 to be used as a feature vectors for a classification process. In the classification process, it was performed by Artificial Neural Network (ANN) with three fully-connected layers, and trained for 500 epochs. In order to evaluate the performance, the 10-flod cross-validation was applied to the networks. The result is that the CNN based bottleneck feature extraction performs a better performance than DNN based which are 99.54% and 98.91% respectively.