【摘要】 雾霾的非均匀随机分布是图像去雾面临的主要挑战性之一。在图像中，雾霾覆盖的范围通常呈现白色或灰白色，降低了该区域的图像信息熵，使得信息在频域内向低频区域聚拢。提出了一种基于分频特征交互的非均匀雾图清晰化算法，首先，对图像进行频域转换，实现多级尺寸压缩和高低频分离。其次，在雾霾分布较高的低频分量，利用Transformer注意力关注机制和全局特征提取能力，增强随机雾霾分布和浓度变化的表征。在高频分量，构建深度差分高频特征增强模块，利用图像自身梯度信息引导，增强图像的边缘细节特征。最后，设计特征交互模块，在Transformer提取到的低频雾霾特征权重指导下，对不同位置和浓度的雾图进行自适应复原，同时实现低频特征与高频特征的层级间信息融合。在4个非均匀雾图数据集上的实验结果表明，所提算法在主观和客观评价均取得优异的效果。更多还原

【Abstract】 The non-uniform and random distribution of haze presents a significant challenge in image dehazing. Haze covered regions often appear white or grayish, reducing the information entropy in these areas and causing data to concentrate in the low frequency domain. This paper proposes a frequency separated feature interaction algorithm for non-uniform image dehazing. First, the algorithm applies frequency domain transformation to achieve multi-level size compression and separation into high and low frequency components. For the low frequency component, a Transformer global feature extraction module is designed to enhance the representation of the random haze distribution and variations in concentration. For the high frequency component, a deep difference enhancement module is developed, leveraging image gradient information to restore and amplify edge details. Finally, a feature interaction module is introduced, where the low frequency haze features extracted by the Transformer guide adaptive restoration across different haze regions and concentrations. This module also facilitates hierarchical interaction between low and high frequency features. Experimental results on four non-uniform haze image datasets demonstrate that the proposed algorithm achieves superior performance in both subjective and objective evaluations.更多还原