Eye state detection based on Weight Binarization Convolution Neural Network and Transfer Learning

Abstract

Detection of eye state can assist the related work in the field of computer vision such as face recognition, expression recognition, pose estimation and human–computer interaction. This paper proposes an Weight Binarization Convolution Neural Network and Transfer Learning (WBCNNTL) based eye state detection method, in which the WBCNNTL is composed of deep convolution neural network and the weight is binarized. The human eye state features can be extracted by Convolutional Neural Network (CNN) effectively, and binary network not only speeds up the computation, but also helps to reduce the storage space and fewer parameters of the model. Transfer learning applies the knowledge or patterns learned from the source domain to different but related fields or problems, which improves the training efficiency of the new model. Experiments on eye state detection are performed using Closed Eyes in the wild (CEW), FER2013 and Zhejiang University Eyeblink (ZJU) Databases, from which the experiment results show the average accuracy obtained by our method are 97.41% on CEW and are 97.15% on ZJU, the computing speed of binary network is faster than non-binary network. Moreover, our method requires less storage space due to lightweight binary model, which maintains better detection capability on CEW compared with some state-of-the-art works.