Abstract
Facial expression recognition (FER) for distinguishing emotional states is challenging due to several issues, including occlusions, extreme lighting conditions, and arbitrary face orientations. To address these challenges, we propose PleFER, a novel framework that leverages plexiform relationship and textural architectonics representations. This model introduces the plexiform relationship mining (PRM), textural architectonics calibration (TAC), and adaptive feature synthesis (AFS) modules. To the best of our knowledge, this study is the first to introduce plexiform relationships for mining architectonics representations in FER tasks. Experimental results on the RAF-DB, KDEF, and RaFD datasets demonstrate that the proposed PleFER model consistently outperforms state-of-the-art methods.