Cascaded deep residual learning network for single image dehazing

Convolutional neural networks (CNNs) have achieved significant success in the field of single image dehazing. However, most existing deep dehazing models are based on atmospheric scattering model, which have high accumulate errors. Thus, Cascaded Deep Residual Learning Network for Single Image Dehazing (CDRLN) with encoder-decoder structure is proposed, which can directly restore the clean image from hazy image. The proposed algorithm consists of a primary network which predicts a residual map based on the entire image, and a sub-network which restores the haze-free image based on the residual image and the original hazy image. The encoder part of CDRLN embeds a context feature extraction module to fuse information effectively. In addition, the two-stage cascaded strategy can avoid feature dilution and restore detailed information, which reduces the color distortion in the dehazing process and generates a more natural, more real and less artifacts dehazed image. Experimental results demonstrate that the CDRLN surpasses previous state-of-the-art single image dehazing methods by a large margin on the synthetic datasets as well as real-world hazy images, and the visual effect of dehazed image is better.