Facial Expression Recognition With Label-Noisy Under Dual-Branch Noise Extraction and Suppression

Noisy labels in Facial Expression Recognition (FER) datasets severely affect the performance of FER models. We propose a novel dual-branch noise extraction and suppression method to address this issue. This algorithm reduces the model’s impact from noisy labels by decreasing the dataset noise ratio and suppressing label-noisy samples. The method comprises three primary stages: sample extraction, pseudo-label generation, and re-training. The approach initially extracts label-noisy samples from the dataset by computing an exponential moving average of the model predictions and the joint probability distribution matrix of noisy and actual labels. The remaining samples form a clean dataset. Next, the training weights of the clean dataset are utilized to assign appropriate pseudo-labels to the label-noisy samples. Subsequently, the noisy labels are replaced with pseudo-labels to create a corrected dataset. The corrected and clean datasets are combined to create the reconstructed dataset, reducing noisy labels within the dataset. Finally, the model is retrained using the reconstructed dataset. Furthermore, this study introduces a novel gradient suppression smoothing function specifically designed to mitigate the impact of label-noisy samples in the dataset during the re-training process. The proposed algorithm is robust, with accuracies of 91.17%, 91.56%, and 91.58% on the RAF-DB dataset with 10%, 20%, and 30% noisy labels, and accuracies of 89.91%, 90.17%, and 89.69% on the corresponding FERPlus.