KTR2DN: Knowledge Transfer with Residual in Residual Dehazing Network
Yihua Lu, Pengfei Phil Duan, Xiongbo Lu, Lei Zhou, Shengwu Xiong · 2023
Single-image dehazing is an essential but challenging computer vision problem. Due to the lack of nonhomogeneous haze datasets, most existing image dehazing methods are only applicable to homogeneous rather than nonhomogeneous dehazing tasks. In addition, the results of existing dehazing methods are always blurred in detail. Thus, a novel network structure, Knowledge Transfer with Residual in Residual Dehazing Network, KTR2DN, is proposed, which consists of two parts: a knowledge transfer network and a super-resolution network using the Residual in Residual (R2) block. The former aims to solve the problem of lacking nonhomogeneous haze datasets and it includes a teacher network and a student network. The teacher network restores abundant clear images so that it can lead the student network to extract more free-haze features. The student network reconstructs the dehazing image from the hazy image through the free-haze features with only a minimal dataset. The latter contains a feature fusion block and a R2 block so that it can fine-tune the results of the student network. The feature fusion network receives both high- and low-frequency feature maps and generates fused feature maps. The R2 block then extracts more semantic features from the fused feature maps. Our method achieves better performance on PSNR and SSIM compared with the state-of-the-art methods.