An Improved Feature Pyramid based Two-Stage Haze Removal Network for Marine Ships

The marine ship image is affected by special weather such as sea fog, which leads to low image contrast and color distortion. In this paper, RDFP-Net which is a modified feature pyramid based two-stage dehazing network is proposed for marine ship images. Considering the loss of details and color distortion of the original image caused by deeper features, a Feature Pyramid Network (FPN) is introduced to retain detailed multi-scale information while a Residual Dense Block (RDB) is combined further to better extract detail feature maps for haze image in the first stage. In addition, the Channel Attention Mechanism (CAM) is employed in the fusion of low-level and high-level feature maps. Thus, more attention is focused on the representative channels by assigning different weight. Finally, more image details are recovered through the second stage. Experimental results demonstrated that RDFP-Net achieves better qualitative visual effects and quantitative indicators with other the state-of-the-art methods.