UCLD-Net: Decoupling Network via Unsupervised Contrastive Learning for Image Dehazing.

From traditional algorithms based on handcrafted prior to learning algorithms based on neural networks, the image dehazing technique has gone through great development. The handcrafted prior-based methods need to first estimate the transmission map and atmosphere light in the atmospheric scattering model separately, and then calculate the final haze-free image, which often leads to a gradual accumulation of errors. In contrast, in the end-to-end neural network-based methods, supervised learning with labels is a major element for the improvement of the dehazing effect. But in the physical situation, paired (hazy, haze-free) images are difficult to collect, which limits the application scope of supervised dehazing. To address this deficiency, we propose a Decoupling Network for image dehazing via Unsupervised Contrastive Learning mechanism which is widely used in self-supervised representation learning, named UCLD-Net. Specifically, we use the estimated transmission map and atmosphere light to design the structure of UCLD-Net and introduce prior knowledge to construct its loss function. It is demonstrated by the experiments that UCLD-Net achieves comparable results in the dehazing experiments on the benchmark RESIDE dataset, which sufficiently verifies its effectiveness.