|
Deep learning has emerged as a pivotal tool in the domain of image dehazing. Deep learning models can learn the mapping between foggy and haze-free images from extensive datasets, thereby achieving the dehazing effect. However, in contemporary deep learning-based dehazing algorithms, the presence of low-quality images engenders challenges, giving rise to issues such as spatial distortions during the dehazing process. Consequently, deep learning-based image dehazing has evolved into a formidable challenge within the realm of computer vision. In order to address this issue, we propose a two-stage conceptual network termed JDN. In the first stage, we employ the image restoration algorithm BOPBL for defect rectification. Subsequently, in the second stage, we employ the non-paired dehazing algorithm RefineDnet. Through experimental validation, we ascertain a significant enhancement in dehazing efficacy. This enhancement can be attributed to the propensity of dehazing algorithms to yield erroneous mappings, there by culminating in distortions. When incorporating image restoration algorithms as a preprocessing step, the acquisition of mappings becomes more amenable, consequently leading to superior dehazing effects. The introduced JDN network is capable of end-to-end training. Moreover, we demonstrate the efficacy of this approach in both image dehazing and quality enhancement. Experimental results on the SOTS-Outdoor dataset reveal that, in comparison to paired algorithms such as FFANet, hardGAN, and PSD, as well as unpaired algorithms including DCP, CycleDehaze, YOLY, and RefineDNet, the JDN exhibits superior performance.
|
