EDGE-COMPUTING CONVOLUTIONAL NEURAL NETWORK WITH HOMOGRAPHY-AUGMENTED DATA FOR FACIAL EMOTION RECOGNITION

Abstract

We design a convolutional neural network with homography-augmented data to deal with facial emotion recognition applications. Different to other convolutional neural networks, our AsicNet is well-designed for embedded CPU and even aiming for ASIC, such as Intel Movidius VPU. We adjust the architecture of our AsicNet and train our AsicNet on the GPU server, meanwhile we consider the computation costs of embedded systems and ASICs. Moreover, we reconsider the deep learning flow and train the homography-augmented data so as to reach higher accuracy. Experimental results on both FER2013 anf JAFFE face datasets show that our AsicNet can not only have high accuracy (72:42% on FER2013; 99:82% on JAFFE) as compared to the state-of-arts but also reach 41.22 millisecond (24.26 FPS) on the embedded CPU and 15.25 millisecond (65.57 FPS) on Intel Movidius VPU to tell the facial emotion from a face image.