An Efficient Dehazing Method Using Pixel Unshuffle and Color Correction

Hongyuan Jing, Kaiyan Wang, Zhiwei Zhu, Aidong Chen, Chen Hong, Mengmeng Zhang

Abstract: Highlights•This paper proposes a new dehazing network with color correction based on deep learning, which does not rely on traditional atmospheric scattering models, but instead uses an end-to-end image transformation approach. Based on the complementary relationship between channels and pixels, the network adaptively adjusts the weights on pixels and channels to enhance the network's ability to represent features.•We present LCA, which expands the number of channels in an image using pixel unshuffle. It then employs channel shuffle to enhance interaction between channels, thereby better integrating multi-channel information and further improving the recovery of local details.•This paper introduces a color loss function to restore the colors of the image, making the restored image more consistent with human subjective perception.•We employ bilinear interpolation to resize feature maps of varying dimensions and superimpose them, which not only makes the weights of attention continuous, but also enables the network to more effectively focus on important information.