RETRACTED ARTICLE: Affective computing model for natural interaction based on large-scale self-built dataset

Abstract

Because of the development of metaverse, robotics, emotional computing has become a very active research topic. And the most important part of affective computing is facial expression recognition, The existing deep learning-based face expression recognition methods are trained on small face expression datasets, and the current datasets have the problems of low quantity, low quality and category imbalance, and the models are prone to overfitting when trained on these small datasets. Compared with other methods using small datasets, this paper proposes a method to build a large face expression dataset by collecting data from the Internet. It also uses both downsampling and data augmentation to solve the category imbalance problem. Meanwhile, this paper experiments the effect of different network structures on face expression recognition based on this brand new dataset. Compared to approaches using small datasets such as raf-db, fer2013, and relatively large but category-unbalanced datasets such as ExpW and AffectNet, the densenet-based model proposed in this paper, trained on this dataset, achieved 97.5% accuracy on the CK+ benchmark, achieve the state-of-the-art performance, which confirms the effectiveness of our proposed approach.