Facial Expression Recognition Based On Feature Learning In Frequency Domain

Facial expression recognition(FER)has become a newly-emerging topic in recent decades,which has important value in the fields of psychology and computer vision.Compared with tra-ditional machine learning,deep learning has superior performance for feature extraction.For deep learning based FER,convolutional neural network(CNN)is widely used.However,mod-els that have superior performance are usually composed of complex convolutional structures,which involve huge computation.On the other hand,it has advantages to process image in frequency domain.However,the existing FER researches that relate to frequency domain are limited to handcrafted feature.Therefore,this paper combines the advantages of deep learning and frequency transform,and presents a network architecture for feature learning in frequency domain,aiming at achiev-ing promising FER performance with lightweight calculation.The innovations of this paper are as follows:1)We propose feature learning in frequency domain.Specifically,we first propose the learnable filter kernel that is used to construct feature extraction layer.Then,we propose the summarization layer following the feature extraction layers to further yield high-level feature.2)We propose the Basic-Fre Net.According to the energy compaction property of discrete cosine transform(DCT),we utilize feature extraction layer and summarization layer to construct the Basic-Fre Net,and realize the first FER algorithm of this paper.3)We propose the DFN-Fre Net.According to the different properties of Basic-Fre Net and CNN,we design feature fusion and construct DFN-Fre Net,which is the second FER algorithm of this paper.4)We propose the Block-Fre Net.Through the analysis of the previous two algorithms,we propose the third FER algorithm: Block-Fre Net,in which the learnable filter kernel is enhanced and dimension reduction is designed.The experimental results show that the proposed feature learning in frequency domain is effective,and the proposed Fre Nets can achieve promising result for FER.Specifically,the Basic-Fre Net can learn discriminative feature on the widely used DCT feature,and achieves higher recognition accuracy for FER.For DFN-Fre Net,the combination of Basic-Fre Net and CNN can further improve FER performance.Finally,the Block-Fre Net can obtain promising result with lightweight calculation.