No Reference Stereoscopic Video Quality Assessment Considering Self-Attention and Different Resolution Level.

In this paper, we propose a no reference stereoscopic video quality assessment (SVQA) method that takes self-attention fusion and different resolution level into account, which reduces loss of detail information. Firstly, considering that left and right branches have semantically inconsistent information, we build a self-attention feature fusion (SAFF) module, which maintains high resolution in both channel and spatial dimensions in the feature fusion process. Secondly, we design a multi-scale module with asymmetric convolution (MSMAC) to obtain multi-scale information and to enhance representation ability of square convolution. Finally, we fuse and weight the features of different resolution levels and then get the final quality score with regressing weighted features. We also consider the depth information between left and right videos and construct the disparity branch. Experimental results on two public stereoscopic video quality databases show that our proposed method outperforms other methods.