RAF-DB、AffectNet、FERPlus

We selected three public datasets to validate the work of this paper, CFMRF_Net. The existing facial expression recognition methods often fail to adequately model the local features of the facial key areas, resulting in limited performance when dealing with complex expressions that have high inter-class similarity and large intra-class differences. To enhance the model's ability to capture the key areas of the face, a cross-fusion multi-level receptive field network (CFMRFN) for facial expression recognition is proposed. The proposed method was experimentally verified on three public datasets: RAF-DB, AffectNet (7cls/8cls), and FERPlus. The overall accuracy rates reached 92.14%, 67.35% (7cls), 63.44% (8cls), and 91.67%, respectively. The dataset is introduced as follows: RAF-DB (Li et al., 2017) is a real-world facial expression recognition dataset that contains 29,672 diverse human face images, covering variations in age, gender, race, posture, lighting, and occlusion. Its basic emotion subset consists of 15,339 images, divided into a training set of 12,271 images and a test set of 3,068 images, with seven basic emotions labeled: happiness, surprise, sadness, anger, disgust, fear, and neutrality. AffectNet (Ali Mollahossein et al., 2017) is currently the largest facial expression dataset, containing over 1 million face images, of which approximately 450,000 have 11 types of emotional labels assigned manually. In this study, a 7-class and 8-class classification setup was adopted. The 7-class setup: 287,651 images were used for training, with 7 classes representing basic emotions; the 8-class setup: added contempt to the original 7 classes. The training set has a class imbalance, while the test set is a strictly balanced set of 500 images per class (totaling 4,000 images). FERPlus (Emad Barsou et al., 2016) is an enhanced version of the FER2013 dataset, containing 35,887 grayscale face images of 48×48 size (28,709 for training, 3,589 for validation, and 3,589 for testing). The labels were re-annotated by multiple annotators, covering 8 emotions: happiness, surprise, sadness, anger, disgust, fear, neutrality, and contempt. The evaluation metric is the overall accuracy of the test set.