基于注意力机制的单图像去雾算法

Single Image Defogging Algorithm Based on Attention Mechanism

图像去雾工作目前还存在真实数据集过少、局部对比度失衡和去雾图像失真等问题。针对去雾图像失真这一问题,提出一种新型去雾网络模型(Densely Resnet with SKattention-Dehaze Net,DRS-Dehaze Net)。首先带雾图像经预处理模块转换为多角度特征输入图,然后设计密集残差架构并引入注意力机制完成特征信息的提取与再分配,最后将特征融合,输出无雾图像。实验结果表明,所提算法相比于其他对比算法有着较好的去雾效果,有效改善了去雾图像失真问题,且在一定程度上对图像的清晰度进行了提升。

Present image-defogging methods have a range of problems:insufficient numbers of real datasets,local contrast imbalance,and defogging image distortion.This paper proposes a novel defogging network model(Densely Resnet with SKattention-Dehaze Net,DRS-Dehaze Net) that mitigates defogging image distortion.First,the fogged image is transformed into a multi-angle feature input map by the preprocessing module.The feature information is then extracted and redistributed through a dense residual architecture with an attention mechanism.Finally,the features are fused to output a fog-free image.Experimental comparison results confirmed a better defogging effect of the proposed algorithm than that of other algorithms.Our model effectively improves the distortion in defogged images and enhances the image clarity to a certain extent.