No-reference stereoscopic image quality assessment (NR-SIQA) predicts stereosc opic image quality without original information,which has been a hot and difficult topic in stereo video system.In this paper,by simulating the binocular vision behavior in human vision system (HVS),a no-ref erence stereoscopic image quality assessment metric is proposed based on binocular feature combination.In the pro posed method,we firstly analyze the binocular combination behavior in NR-SIQA,and propose a binocular combinat ion mode for quality prediction.Then,in order to represent local and global prop erties of stereoscopic images,the features extracted by learning and statistic analysis are combined together as the final perceptual features.Finall y,a regression model is learnt by machine learning algorithm to map human subjective scores and features,and the learnt regression model and the binocular combination model are utilized to predict the quality of a test stereo scopic image pair.Experimental results demonstrate that by applying the proposed model on three public database s,the Pearson linear correlation coefficient (PLCC) and Spearman rank order correlation coefficient (SROCC) of th e proposed method are higher than 0.93in symmetric database and 0.87in asymmetric databases.Compared with the state-of-art SIQA methods, the proposed one outperforms most of them,which indicates that the metric is fa irly good and can predict human visual perception very well.