Existing image dehazing methods suffer from problems of insufficient dehazing, distortion, and low color contrast. Aiming at this problem, a deep learning single-image dehazing method based on multi-scale segmentation is proposed in this paper. The study found that the haze information in the haze image will decrease with the increase of frequency. Therefore, the haze image is first decomposed into four sub-images of different frequency domains through image segmentation in this article. A dehazing network model composed of four sub-network channels with different complexity is then constructed to extract the haze information contained in each sub-image. After the transmission sub-images are generated, the image fusion technology is used to obtain the final transmittance map. Finally, the haze-free image is obtained based on the physical model of atmospheric scattering. Experimental results on the synthetic and real images dataset show that the proposed method achieves significant dehazing effect and high color contrast with no distortion, showing superior performance than other dehazing methods.

现有的图像去雾方法存在去雾不足、失真和颜色对比度低的问题。针对这一问题，本文提出了一种基于多尺度分割的深度学习单幅图像去雾方法。研究发现，雾霾图像中的雾霾信息会随着频率的增加而减少。因此，本文首先通过图像分割将雾霾图像分解为四个不同频域的子图像。然后构建由四个不同复杂度的子网络通道组成的去雾网络模型，以提取每个子图像中包含的雾度信息。透射子图像生成后，利用图像融合技术得到最终的透射图。最后，基于大气散射物理模型得到无雾图像。在合成图像和真实图像数据集上的实验结果表明，该方法实现了显着的去雾效果和高色彩对比度且无失真，表现出优于其他去雾方法的性能。