Transfer Learning: An Enhanced Feature Extraction Techniques for Facial Emotion Classification

The process of extracting features from images from the scratch in the field of machine learning is a very challenging and time-consuming task. Previous work focused on a technique for feature extraction in which a CNN is pre-trained on some data sets and then in turn used to learn the pattern from other datasets on which it was not originally trained. This paper proposed transfer learning as the ability of a pre-trained CNN model to learn patterns from data which it was not originally trained on. Several pre-trained CNN architectures have been proposed by researchers to improve the system performance. Some of these include the ResNet, Xception, VGG۱۶, VGG۱۹, InceptionResnetV۲, InceptionV۳, DenseNet, MobileNet, etc., all trained on ImageNet datasets. However, not all of these networks have the ability to effectively perform feature extraction on facial datasets. This paper attempts to compare the efficiencies of five pre-trained networks namely, VGG۱۹, VGG۱۶, ResNet۵۰, inceptionResNetV۳, and InceptionV۲ which have been pre-trained on the ImageNet dataset as features extractors on the DISFA plus facial emotion dataset for classification. The extracted features are used for training and testing a simple linear regression classifier. Testing the classifier network with features extracted by these networks produced different accuracies. It was discovered in the process that out of the five networks being tested, the VGG۱۹ architecture performed more accurately than other pre-trained networks on the classifier with an accuracy of ۹۷%. The work, therefore, presumes that the VGG۱۹ network has a better generalization ability on facial data than other networks.

Keywords:

Convolutional Neural Network , Machine Learning , Transfer Learning , Pre-trained network