Reducing Computational Requirements of Image Dehazing Using Super-Resolution Networks
Shyang-En Weng, Y. Ye, Ying-Cheng Lin, Shaou-Gang Miaou · 2023
The task of image dehazing (DH) is an indispensable part of Advanced Driver Assistance Systems. This study proposes a method that combines image super-resolution (SR) and DH, focusing on how the models for DH and SR can be combined to reduce the overall computational cost. The considered DH networks include AOD-Net and GridDehazeNet, and the considered SR methods include bicubic interpolation, SRCNN, FSRCNN, and a modified version of SRCNN proposed in this study to help reducing the overall DH computations. In addition, a joint network training strategy for DH/SR combination is proposed. The experimental results show that the proposed method can reduce the computational complexity by about 70% compared to the original DH network – GridDehazeNet, while effectively preserving the image quality and clarity.