Pyramid feature boosted network for single image dehazing

In this paper, a Pyramid Feature Boosted Network is proposed for single image dehazing, which leverages the encoder-decoder structure and benefits from two core modules to achieve high-quality image recovery. Since image detail loss is a common problem in image restoration, we design a Feature Boosted module based on the Strengthen-Operate-Subtract boosting strategy to increase the quality of the image. This module innovatively incorporates multi-scale latent features to replenish the lost signals. In addition, to release the heterogeneous haze, a novel Mixture Attention unit is proposed to reinforce the important information in multiple dimensions and highlight the main object in the image from background. Extensive evaluations and simulation results show the proposed methods outperform the State-Of-The-Art (SOTA) methods on both synthetic datasets and real-world images.