Hybrid architectures based on Deep-wavelet networks for emotion recognition as a tool to support personalized therapies

Human emotions influence social interaction between individuals and serve as a foundation for effective communication. For healthcare, tools based on emotion recognition can be extremely useful in therapeutic contexts, especially for patients who have difficulty expressing emotions, often unable to express the discomfort associated with treatments. Aiming to contribute to this scenario, this work presents a comparison between hybrid architectures, composed of a Deep-Wavelet Neural Network (DWNN) and a Random Forest. For training and testing the architectures, the FER2013 database was used. As a result, our best architecture, composed of a 4-layer DWNN and Random Forest, obtained 70.77% of accuracy in the training stage and 81.21% in the testing stage, using the balanced dataset and feature selection. This performance was achieved with only approximately 34.06% of the DWNN attributes, equivalent to 436 attributes out of 1,280. The intention is that the proposed architecture composes a system for recognizing emotions through facial expressions for personalizing therapies.