Abstract: Compared with the planar image quality assessment, the stereoscopic image quality assessment is more challenging because it needs to consider the complex characteristics of binocular fusion of left and right views and binocular interaction. For this reason, this paper proposes a no-reference stereoscopic image quality assessment network based on multi-task learning by combining the left and right visual perception characteristics of the human eye. Firstly, a two-channel convolutional neural network is built for feature extraction and quality assessment tasks of the left and right views. Secondly, an adaptive binocular interaction module is designed to simulate binocular interaction mechanism in HVS visual cortex. Finally, considering the characteristics of binocular fusion and rivalry of the human eyes, the fusion and disparity images of the left and right view feature blocks are connected with monocular extraction features as the stereoscopic features to evaluate the stereo image quality. The method on the Waterloo-P1 and Waterloo-P2 public databases exhibits that the Pearson linear correlation coefficients are 0.976 and 0.982 respectively, and Spearman rank correlation coefficients are 0.974 and 0.980 respectively, which is better than some other state-of-art algorithms. The experimental results show that compared with the single-task model, the proposed method achieves better performance, showing a high consistency with human subjective evaluation. In addition, the cross-dataset experiments also validate the generalization capability of the multi-task learning algorithm, which has good universality.