A performance comparison of two emotion-recognition implementations using OpenCV and Cognitive Services API

Abstract

Abstract.Emotions represent feelings about people in several situations. Various machine learning algorithms have been developed for emotion detection in a multimedia element, such as an image or a video. These techniques can be measured by comparing their accuracy with a given dataset in order to determine which algorithm can be selected among others. This paper deals with the comparison of two implementations of emotion recognition in faces, each implemented with specific technology. OpenCV is an open-source library of functions and packages mostly used for computer-vision analysis and applications. Cognitive services is a set of APIs containing artificial intelligence algorithms for computer-vision, speech, knowledge, and language processing. Two Android mobile applications were developed in order to test the performance between an OpenCV algorithm for emotion recognition and an implementation of Emotion cognitive service. For this research, one thousand tests were carried out per experiment. Our findings show that the OpenCV implementation got a better performance than the Cognitive services application. In both cases, performance can be improved by increasing the sample size per emotion during the training step.

Emotion Recognition using AutoEncoders and Convolutional Neural Networks

“…Since accuracy involves both specificity and sensitivity in its calculations, the accuracy for these five emotions are close to 90% as well. In a previous research [30], our findings were no higher than 80% for all emotions accuracy. Another improvement was detected for the disgust feeling, which presents an approximate 85% value in both accuracy and precision.…”