Facial Expressions Recognition by Structuring the Embeddings Space
Mihai Boeru, Andrei Racoviţeanu, Corneliu Florea · 2021 International Conference on e-Health and Bioengineering (EHB) · 2021
Lately, Facial Expression Recognition (FER) gained in popularity in the medical field due to their importance in understanding the human psychology or detecting various pathologies. In this paper, we propose a more efficient solution for recognizing facial expressions using a more convenient structuring of the data. To accomplish this aspect, a clustering loss entitled Large Margin was used in addition to the classic Cross Entropy, in order to improve discrimination in the embeddings space. The main study consisted of trying different distance metrics (L1, L2, cosine similarity) between embeddings, to examine which of them offers a better structuring of the data. The solution is tested on two benchmarks with very good results.