To assess stereoscopic image quality with various types of distortion effectively,a no-reference (NR) stereoscopic image quality assessment is propos ed which uses extreme learning machine (ELM) with kernel to learn the features based on image quaternion wavelet transform (QWT).Firstly,i t adopts the left view image and the right view image of stereoscopic image as the input of the structural simila rity (SSIM) based dense stereo matching algorithm to obtain 3D perceptual maps of the stereoscopic image:an estimated disparity map, an estimated disparity confidence map and an estimated disparity-compensated right view image.Secondl y,it processes the left view image,the right view image,the estimated disparity map,the estimated disparit y-compensated right view image with QWT.Thirdly,it computes the energy and weight standard deviation of all i mage QWT coefficient.Then it computes the statistical feature entropy and median of the estimated disparity c onfidence map.At last,these features above are used as the input of ELM with kernel to predict the quality o f the tested stereo images.The experiment is based on no-cross image material in LIVE 3D image quality databas e,and the Spearman rank ordered correlation coefficients (SROCCs) are 0.926in phase I a nd 0.914in phase II,which indicates proposed method is consistent with subj ect quality assesment.