Dehazing Network Based on U-Net Structure and Residual Block

With the development of deep learning technology, the dehazing method based on convolutional neural network has also developed rapidly. However, it still faces some problems such as incomplete dehazing and difficulty in detail restoration. Aiming at the problems, we propose a dehazing network based on U-Net structure and residual module. In the down sampling process, we introduce the residual block of attention mechanism, which effectively improves the feature extraction and expression ability of the network. In the middle part of the network, the smooth dilation convolution residual block is used to expand the receptive field of the network and improve the restoration effect of small items. At the same time, we introduce a feature fusion mechanism based on attention mechanism, which can reduce the loss of information and weight the down-sampling features into intermediate features. In the process of up sampling, the method of adding elements instead of feature stitching is used to reduce the number of parameters and realize the function of jump connection. Finally, dense convolution residual blocks are added to the jump connection between input and output to further improve the performance. The experimental results show that the PSNR of this method can reach 32.893, which is better than the existing methods.