Based on the minimum energy error, the disparity image of the left and right views is obtained, and the steerable pyramid decomposition which has four scales and twelve orientations are carried on the left, right views and their disparity image. Each image can get one high frequency sub-band and fourty-eight directional sub-bands. The bivariate generalized Gaussian distribution fitting is carried on fourty-eight pairs of directional sub-bands corresponding to the decomposition of left and right views. Their shape and scale parameters are extracted, and across scales, spatial correlation statistical features from all orientation sub-bands are extracted. These features are input to support vector regression (SVR) to train and predict the stereo image quality score. Results show that the performance evaluation index of the quality evaluation model in LIVE 3D database, Spearman rank order correlation coefficient (SROCC), linear correlation coefficient (CC) are all above 0.93, which has good consistency with the subjective evaluation of human.