Facial expression recognition using frequency multiplication network with uniform rectangular features

Facial expression recognition (FER) is a popular research field in cognitive interaction systems and artificial intelligence. Many deep learning methods achieve outstanding performances at the expense of enormous computation workload. Limiting their application in small devices or offline scenarios. To cope with this drawback, this paper proposes the Frequency Multiplication Network (FMN), a deep learning method operating in the frequency domain that significantly reduces network capacity and computation workload. By taking advantage of the frequency domain conversion, this novel deep learning method utilizes multiplication layers for effective feature extraction. In conjunction with the Uniform Rectangular Features (URF), our method further improves the performance and reduces the training effort. On three publicly available datasets (CK+, Oulu, and MMI), our method achieves substantial improvements in comparison to popular approaches.