Fast Video Facial Expression Recognition by Deeply Tensor-compressed LSTM Neural Network for Mobile Device

Poster: Mobile devices usually suffer from limited computation and storageresource which seriously hinders them from deep neural network applica-tions. In this paper, we introduce a deeply tensor-compressed LSTM neuralnetwork for fast facial expression recognition (FER) in videos on mobiledevices. Firstly, a spatio-temporal FER LSTM model is built by extractingtime-series feature maps from facial clips. The LSTM model is further deeplycompressed with tensorization. Based on dataset of Acted Facial Expressionin Wild (AFEW) 7.0, experimental results show that the proposed methodachieves 55.60% classification accuracy; and significantly compresses the sizeof network model by 219×. Our work is further implemented on RK3399ProIoT device with Neural Process Engine, and the runtime of feature extractionpart can be reduced by 12.83×with only 7.73W power consumption.