Comparative Evaluation of Facial Emotion Recognition Models for Customer Satisfaction Analysis

Using artificial intelligence (AI) tools to identify emotions is beneficial for evaluating customer satisfaction after interacting with a product or service. This paper aims to conduct a comparative analysis between two facial expression detection solutions. The paper focuses on analyzing the DeepFace and Azure Custom Vision services for a set of processed image data from a public database. The aim of the research is to determine the extent to which these systems can be integrated into real-time customer satisfaction monitoring applications. This approach is superior to subjective assessments because it is unmediated by words or a masked formulation of the true emotion. The analysis was performed on a limited number of images per emotion category due to cost constraints. The number of tests was maintained the same for the two services to produce an equidistant analysis. The categories of emotions analyzed were anger, disgust, fear, happiness, sadness, surprise, and neutrality. The performance of each instrument was evaluated by the accuracy identified using confusion matrices. The results showed superior performance for the DeepFace service, which achieved an accuracy of 87.29%, compared to Azure Custom Vision, which only achieved 82.99%. These results were obtained for the experimental dataset used. The research shows the importance of the dataset used in training AI models. In this case, the 48x48 pixel image quality compromises the model's ability to correctly recognize users' emotional states. A second experimental test was made by increasing the number of images to 28,709 images for the DeepFace service with the objective of analyzing the accuracy improvement. A slow increase was observed from 87.29% to 89.84%. The findings highlight that DeepFace is superior to Azure Custom Vision, but in this version, it cannot be used for customer satisfaction monitoring due to its low performance. The research will continue using higher-quality datasets and will be reapplied to an extended class of instruments in future research.