Dehazing CNNs Loss Functions Analysis

This paper examines the performance of several loss functions in the context of image dehazing. We considered two CNN architectures for dehazing (AOD-Net and UVM-Net), trained on nine loss functions: L2, L1, Smooth L1, Huber, PSNR, SSIM, MS-SSIM, HaarPSI, and PSNR-HVS. The latter two functions have not been previously used as loss functions, only for approximating human perceptions in various image processing tasks. The resulting 18 CNN models were tested on three different image dehazing datasets. Our results demonstrate that the perceptual loss functions (PSNR-HVS and HaarPSI) outperform the traditional loss functions for the CNN image dehazing models considered.