Blind Dehazed Image Quality Assessment: A Deep CNN-Based Approach

Research on image dehazing has made the need for a suitable dehazed quality assessment (DIQA) method even more urgent. The performance of existing DIQA methods heavily relies handcrafted haze-related features. Since hazy images with uneven haze density distributions will result in after dehazing, manually extracted feature expression is neither accurate nor robust. In this paper, we design deep CNN-based without requirement. Specifically, propose blind model (BDQM), which consists three components: preprocessing, extraction network (HFNet), and an improved regression (IRNet). HFNet, perceptual information enhancement (PIE) module to learn powerful representations enhance capability according channel attention, multiscale convolution residual concatenation. IRNet aims aggregate all patch prediction whole image, where effect inhomogeneous distortion from procedure attenuated via specifically designed attention (PA) mechanism. Experimental results benchmark datasets demonstrate effectiveness superiority proposed architecture over state-of-the-art methods.