A Comparative Study of Image Dehazing Based on Attention-Net and U-Net Atmospheric Light Estimation
Yuehong Cui, Jinghua Wu, Yuwei Chen, Chunling Cai · 2020
The effect of image dehazing depends largely on the estimation effect of atmospheric light. In this paper, the Attention-Unet network is introduced to estimate atmospheric light. The model trained by the attention gate mechanism can suppress the irrelevant area of atmospheric light in the input image through implicit learning, and highlight the significant characteristics of the area that has a large impact on atmospheric light, so as to solve the problem of parameter invariability caused by U-Net's difficulty in self-learning. In this paper, the pyramid dense network estimation network is used to estimate the transmission diagram, and then the attention U-net is used to predict atmospheric light. The results of the two networks are obtained by atmospheric scattering model. Meanwhile, the output transmission map and output figure to haze through joint discriminator, can better use the structure of the correlation between them. The experimental results show that the effect of dehaze is better than that of DCPDN by using UNET to estimate atmospheric light, and the performance of SSIM and PSNR is improved on the same data set NNA-depTH2.