SDCNet:Spatially-Adaptive Deformable Convolution Networks for HR NonHomogeneous Dehazing

In recent years, the field of image dehazing has garnered increasing attention. Many deep learning models have demonstrated exceptional capabilities in removing homogeneous haze, yet they often perform suboptimally when faced with the challenge of non-homogeneous dehazing. One of the primary issues is that these models are trained under conditions of homogeneous haze, which does not align with the characteristics of real-world haze scenarios. non-homogeneous haze typically leads to structural distortion and color shifts in images. Another contributing factor is the limited scale of datasets available for non-homogeneous dehazing, which hampers the training of robust models. To address these challenges, we have designed a Spatially-Adaptive Deformable Convolution Networks (SDCNet). The first branch of our model incorporates a high-level prior model that serves as an encoder for extracting high-level features from the image. The second branch is composed of a lightweight network specifically tailored to extract low-level features from hazy images. Our model fuses the information from both branches and combines progressive training as well as dynamic data augmentation strategies to obtain visually pleasing dehaze results. Extensive ablation studies have been conducted, substantiating the effectiveness and feasibility of our proposed methodology. Furthermore, in the NTIRE 2024 Dense and NonHomogeneous Dehazing Challenge, we achieved the best performance in terms of PSNR, SSIM, and MOS.

1. Introduction

Our results on the NTIRE 2024 Dense and NonHomogeneous Dehazing Challenge, achieving the best performance in terms of