扩散模型与傅里叶滤波器结合的图像去雾方法

Image dehazing method combining diffusion model and Fourier filter

当前图像去雾算法研究主要聚焦于构建结构精巧的神经网络,利用扩散模型以生成式方法进行图像去雾的研究比较匮乏,且扩散模型与频域分析结合的研究也相对较少,这导致许多算法在面对真实有雾图像时去雾效果不佳。针对上述问题,本文基于分数匹配扩散模型IR-SDE提出了一种图像去雾方法Firs-Net。该方法利用扩散模型强大的图像生成能力,通过迭代式的方式从有雾图像逐渐恢复清晰无雾图像。Firs-Net中引入了傅里叶特征融合模块,该模块可在不增加参数量且无需微调神经网络的前提下,帮助扩散模型更好地从频域角度分析并融合特征。实验结果表明,Firs-Net在真实有雾数据集上的主观视觉和客观指标都表现优异,其中,在真实非均匀雾气数据集NH-HAZE中的PSNR达到21.91,在真实浓雾数据集Dense-HAZE中的PSNR达到17.40,分别比次优方法领先13.94%和5.52%。

The current research on image dehazing algorithms mainly focuses on building intricately structured neural networks.There is a lack of research on using diffusion models and generative methods for image dehazing,and there is also relatively little research on combining diffusion models with frequency domain analysis.This results in many algorithms having poor dehazing effects when processing real hazy images.In response to the above issues,a Firs-Net image dehazing method is proposed based on the score matching diffusion model IR-SDE.This method utilizes the powerful image generation ability of diffusion models to gradually restore clear and hazy-free images from hazy images in an iterative manner.Firs-Net introduces a Fourier feature fusion module,which can help the diffusion models better analyze and fuse features from a frequency domain perspective without increasing the number of parameters and fine-tuning the neural network.The experimental results show that Firs-Net performs excellently in both subjective visual and objective indicators on real hazy datasets,with a PSNR of 21.91 in the NH-HAZE dataset of real non-uniform haze,and a PSNR of 17.40 in the real dense haze dataset Dense-HAZE,which is 13.94%and 5.52%ahead of the suboptimal method,respectively.