Channel Splitting Network for Single Image Dehazing
Zhiming Su, Lie Wang · 2020
Haze affects the image quality and image analysis, so it is necessary to recover the undamaged content from the blurred image. We propose an end-to-end channel splitting network to directly reconstruct clear and haze-free images without relying on the atmospheric scattering model. By introducing the adjacent connection network, a channel splitting residual 1*1 convolution module and a channel splitting pyramid dilated convolution module are constructed to improve the detailed information of the reconstructed image. We propose an attention mechanism with local features, which further improves the quality of haze-free images. The network that can identify the residual of hazy is constructed by the above modules, which is based on three series connected U-shaped networks with only one downsampling. This network optimization reconstructs the loss function by adaptive weighting. Extensive experiments on RESIDE data set demonstrate that compared with several selected state-of-the-art algorithms, the proposed method achieves significant improvement in quantity and quality.