基于解耦重构的弱监督单幅图像去雾算法

Weakly-supervised single image dehazing algorithm based on decoupling reconstruction

目前,基于深度学习的单幅图像去雾算法依赖于成对的有雾和无雾图像进行训练,在真实场景的去雾效果受限于合成数据集中有雾图像的合成效果。针对成对数据集获取困难的问题,提出了一种基于解耦重构的弱监督单幅图像去雾算法。该算法分别利用内容编码器和雾分布编码器,从输入图像中解耦出图像内容和雾分布信息,最终利用生成器模型对提取的信息进行图像重构。和现有去雾算法相比,所提出的去雾算法在图像色彩还原和雾霾去除效果上较好,还原结果更接近真实的无雾场景,在视觉效果和客观指标上有明显优势。

By far,deep learning-based single-image dehazing algorithms rely on paired haze and haze-free images for training.The dehazing results on real haze images are limited by the facticity of synthetic datasets.Aiming at the difficulty of acquiring paired data sets,the weak-supervised single de-hazing algorithm based on decoupling reconstruction is proposed.The algorithm uses the content encoder and the haze distribution encoder respectively to decouple the image content and fog distribution information from the input image,and uses the generator model to reconstruct the extracted information.Compared with the existing dehazing algorithm,the proposed dehazing algorithm is superior in image color reproduction and haze removal,and the restoration result is closer to the real haze-free scene.Experimental results on both visual effects and objective indicators demonstrate that the proposed model performs favorably against the state-of-the-art methods.