Deep Leaning-Based Facial Expression Recognition in FER2013 Database: An in-Vehicle Application

Abstract

This work presents a deep learning-based approach for the evaluation of facial expression recognition (FER) performance. The main objective is to develop a deep convolutional neural network (CNN) to perform FER using the publicly available benchmark dataset, the FER2013 dataset. The FER2013 dataset includes hand-based facial occlusion, incorrectly cropped or partial images, images with glasses, low-resolution images, etc., which are close to real driving complex scenarios. Two custom CNN models and a pre-trained VGG16 model are evaluated for the FER task. The deep CNN model with 10-layer architecture shows the best performance accuracy of 68.34%. This deep CNN model can be used to monitor driver behavior from front face images captured via dashboard camera and alert the driver to improve their driving style for a safe drive.