Facial Expression Recognition Based on TripletLoss and Attention Mechanism

The paper proposes a facial expression recognition method called Triplet-Loss Attention Network (TAN), which aims to address the problem of large intra-class and inter-class distances in facial expression recognition. The method uses a self-attention mechanism to calculate the weight of triplet expression samples, which helps in determining the influence of key regions on expression recognition. The method uses images with attention weights above a certain threshold as training samples to form new hard triplets. The distance between high-weight samples in the three sets of samples is calculated using the Mahalanobis distance formula, and the difference in distance between high-weight groups of triplets is calculated using the Maharaja Loss function. The Triplet Loss is mainly used as the loss function in TAN, and the model is jointly optimized using Triplet Loss, Mahalanobis Loss, and Cross Entropy Loss functions to improve the performance of the model in facial expression recognition. Experimental results show that TAN performs well in alleviating the intra-class and inter-class distance problem and has good robustness and generalization performance. On the RAF-DB and FERPlus datasets, TAN achieves recognition accuracies of 88.40% and 88.73%, respectively, which is 1.37% and 0.18% higher than the previous state-of-the-art methods.