基于扩散模型与移位窗口策略下的图像去雾方法

Image Dehazing Method Based on Diffusion Model and Shifted Window Strategy

摘要:

针对传统去雾网络存在的去雾后图像细节损失、去雾效果不理想等问题，本文提出一种基于扩散模型与移位窗口策略下的图像去雾方法。首先，使用压缩编码器将有雾图像和对应的无雾图像进行联合特征提取，再使用扩散模型（Diffusion Models, DMs）对特征进行优化。接着将优化后的特征与原始有雾图像一同作为输入，送入一个以U-Net(U-Network)结构为基础结构的去雾网络中，该网络将特征图通过移位窗口分区后，在每一个分区内使用通道注意力和空间注意力并行处理特征图像，能够更精确地捕捉图像中的信息相关性以及增强对重要特征的关注，经典的编码器-解码器结构能够有效地恢复图像的细节和边缘信息。实验结果表明，本文所提出的方法在多个公开数据集上PSNR提升0.47~2.62db。

Abstract:

To address issues such as detail loss and suboptimal dehazing performance in traditional dehazing networks, this paper proposes an image dehazing method based on a diffusion model and a shifted window strategy. First, a compressed encoder is employed to jointly extract features from hazy images and their corresponding haze-free images. The extracted features are then optimized using a diffusion models (DMs). Subsequently, the optimized features, along with the original hazy image, are fed into a dehazing network based on the U-Network (U-Net) structure. This network partitions the feature map into shifted windows and applies channel attention and spatial attention in parallel within each partition. This approach enables the model to more accurately capture correlations within the image and enhance focus on critical features. The classic encoder-decoder structure effectively restores image details and edge information. Experimental results demonstrate that the proposed method improves the PSNR by 0.47 to 2.62 dB across multiple public datasets.