Smart Monitor Health System: Face Expressions Recognition

Abstract

The emerging success of digital social media has had an impact on several fields ranging from science to economy and business. This has been particularly relevant for the marketing industry which centers its activity on digital social interactions between branding and the end-consumer, in order to increase their market competitiveness. Therefore, there is an invested interest in emotion detection and recognition technology from facial expressions. This work is marked by a detailed research about the concepts and the existent methodologies behind the Automatic Facial Expression Recognition (AFER) systems, as well as an evaluation of their effectiveness. Additionally, the most relevant models were tested, in order to discern the most adequate one for facial expressions recognition. A comparison was made between the traditional methodologies and Deep Learning, a recent trend in Pattern Recognition. Both these domains have challenging inner workings. Traditional methods are strongly dependent on the input, thus any transformation in the dataset will influence the model for more adjustments. The Deep Learning methods are more adaptable to variation, however refining their hyperparameters might be an exhaustive work. Here, we explored the value of Deep Learning by focusing on recent technological breakthroughs, particularly with Convolutional Neural Networks (CNN). Incremental steps were made in order to deploy the better solution to the network architecture. After some preliminary experiments using the more recent and complex networks (such as Googlenet, Alexnet) we ended with the Lenet-5 as a baseline. We found that Lenet-5 simplicity was better suited for the system constraints (dataset dimension, faces size and composition). The Cohn-Kanade Extended dataset was chosen for testing our proposed CNN model. In an attempt to improve the results we also have augmented the dataset with random perturbations from a wide set, including: skew, translation, scale, and horizontal flip. Our refinements to the model led to a 90% overall accuracy when taking static images as an input. To further validate our results we built and present a real-time framework. Based on the data collected, the deep model has emerged as a promising approach for AFER systems.