Facial Emotion Recognition Algorithm in the Wild Based on Auxiliary Label Spatial Graph

Facial emotion recognition (FER) plays a crucial role in computer vision, widely used in driver status detection, criminal interrogation, and human-computer interaction. However, the challenging conditions such as illumination variations, occlusions, and blurring, has made accurate annotation difficult and led to a significant number of uncertain labels. To solve this problem, we present an end-to-end network structure called Auxiliary Label Space Graph-based Latent Distribution Mining (AL-LDM). Firstly, by employing a multi-branch network, AL-LDM mines the latent distribution of labels in images, which is more closely with subjective intuition. Subsequently, an auxiliary feature graph is constructed using KGraph to adjust the confidence level of the images. Finally, the method dynamically assigns different weights to facilitate effective learning based on label uncertainty. Experimental results demonstrate the effectiveness of AL-LDM. Compared to the baseline, it achieves an improvement of 2.36% and 3.93% on the RAF-DB and AffectNet datasets, respectively.