Cross-channel Fusion Image Dehazing Network with Feature Attention

Yong Liu, Xiaorong Hou

Abstract: In this paper, we propose a cross-channel fusion image dehazing network with feature attention (CFDN), which directly restores the final clear image from the hazy input. The network design is motivated by three strategies, namely cross-channel fusion, feature attention mechanism, and local residual learning. We show that they are effective for image dehazing problem. By separately extracting the features of the RGB three-channel components of the hazy image, the CFDN can effectively learn the different haze information in different color spaces. We develop two key components in our CFDN: internal and external channel feature attention network (Inter-CFAnet and Exter-CFAnet). Inter-CFAnet focuses on assigning more weight to important feature components; while Exter-CFAnet leverages the weighted fusion to adaptively merge cross-channel complementary features effectively. We also embed pixel feature attention network (PFAnet) to solve the problem of uneven haze distribution haze in the image. In addition, local residual learning allows the network to bypass less useful information through multiple skip connections. Experiments demonstrate that the proposed CFDN performs favorably against previous state-of-the-art image dehazing methods both quantitatively and qualitatively.