Curriculum learning-based convolutional neural network for removing dense haze from a single image
Qingbo Wu, X. W. Zheng · IET conference proceedings. · 2025
Dense haze often severely impairs the quality of a photograph, such as weakened details and shifted colors of scenes. Current dehazing methods tend to remove light-and mediate-density of haze from a single image in a one-shot pipeline. However, these approaches fail in the case of dense haze and usually lead to significant residual haze-style artifacts. For dense haze removal, we propose a new pipeline that treats dense haze as a concatenation of several pieces of light haze and progressively removes them to recover clear scenes. To this end, we introduce a network consisting of multiple U-shaped encoder-decoder subnets and train it using our synthesized different densities of light-haze images based on the curriculum learning. In addition, extensive experiments on the Dense-Haze dataset are carried out to demonstrate that our method performs favourably against four state-of-the-art dehazing approaches in quality and quantity.