Improving Lightweight Convolutional Neural Network for Facial Expression Recognition via Transfer Learning

Abstract

Image-based facial expression recognition is an important problem especially for analyzing the human emotion or feeling under a specific condition, such as while watching a movie scene or playing a computer game. Furthermore, the convolutional neural network (CNN) is one of the underlying technology proven to be applicable to image-based facial expression recognition problem. Unfortunately, the available CNN architecture that applied for image-based facial expression recognition problem only focuses on the accuracy instead of other factors, such as the number of parameters and the execution time. In this paper, we investigated whether transfer learning from a medium-size and large-size dataset is feasible to improve the performance of lightweight CNN architecture on image-based facial expression recognition problem. We use lightweight residual-based CNN architecture originally used for CIFAR dataset to analyze the effect of the transfer learning from five different datasets, including CIFAR10, CIFAR100, ImageNet32, CINC-10, and CASIA-WebFace. The FER+ (Facial Expression Recognition Plus) dataset is used to evaluate the lightweight CNN architecture performance. Experiments show that our lightweight CNN classifier can also be improved even when the transfer learning performing from middle-size dataset comparing when training the classifier from scratch.