Quadratic Polynomial Residual Network for No-Reference Image Quality Assessment

Residual connection has become an essential structure of deep neural networks. In residual connection, shallow features are directly superimposed to deep features without any processing. In this paper, a quadratic polynomial residual module is designed to increase the nonlinear fitting ability of the network. As the name suggests, this module superimposes quadratic polynomials of shallow features onto deep features. In this way, the series of two modules has the fitting ability of a quartic polynomial. The fitting ability of the network increases exponentially with the number of layers. According to Taylor’s theorem, it can be concluded that this module effectively improves the fitting ability of the network. Meanwhile, the image patches containing more information have greater contribution to image quality assessment. The patches are screened according to the two-dimensional information entropy, which reflects the information amount of patches. Based on the above two points, a quadratic polynomial residual network with entropy weighting and multi-receptive field structure is proposed for no-reference image quality assessment. The experimental results show that the proposed algorithm achieves high accuracy and more effectively fits the human visual system.