刘文进, 王鑫, 林森, 陶志勇, 徐光宪. 基于融合残差与注意力机制的图像去雨网络[J]. 云南大学学报(自然科学版), 2022, 44(5): 940-948. doi: 10.7540/j.ynu.20210644
引用本文: 刘文进, 王鑫, 林森, 陶志勇, 徐光宪. 基于融合残差与注意力机制的图像去雨网络[J]. 云南大学学报(自然科学版), 2022, 44(5): 940-948. doi: 10.7540/j.ynu.20210644
LIU Wen-jin, WANG Xin, LIN Sen, TAO Zhi-yong, XU Guang-xian. Image deraining network based on fusing residuals and attention mechanisms[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(5): 940-948. DOI: 10.7540/j.ynu.20210644
Citation: LIU Wen-jin, WANG Xin, LIN Sen, TAO Zhi-yong, XU Guang-xian. Image deraining network based on fusing residuals and attention mechanisms[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(5): 940-948. DOI: 10.7540/j.ynu.20210644

基于融合残差与注意力机制的图像去雨网络

Image deraining network based on fusing residuals and attention mechanisms

  • 摘要: 针对已有去雨网络无法更好地保留图像边缘纹理细节的问题,提出一种基于融合残差与注意力机制的图像去雨网络. 首先,对原始图像进行浅层学习,从中提取出关键的浅层特征;然后,为了避免浅层信息的丢失,将多注意力模块与残差结构融合构成注意力残差块,将已提取的浅层特征信息输入到其中进行更高层语义特征学习,在其中运用多注意力模块获得不同形状、尺寸雨纹的多尺度特征,构建不同通道间的依赖性以及使网络更加关注雨纹和高频区域的图像信息特征;最后,通过卷积层进行特征重构,得到去除雨纹的清晰图像. 实验结果表明,新算法在Rain100H、Rain100L和Rain12公开测试集上分别获得28.91、36.86 dB和35.14 dB的峰值信噪比,以及95.0%、99.0%和97.1%的结构相似性. 新算法客观评价指标均优于其他对比算法,主观视觉效果得以有效提升,能够去除不同密度雨纹的同时更好地保留图像的细节信息.

     

    Abstract: Aiming at the problem that the existing rain removal network cannot better preserve the texture details of the image edge, an image rain removal network based on the fusion residual and attention mechanism is proposed. First, shallow learning is performed on the original image, from which key shallow features are extracted. Then, in order to avoid the loss of shallow information, the multi-attention module is fused with the residual structure to form an attention residual block, and the extracted shallow feature information is input into it for higher semantic feature learning, in which the multi-attention module is used to obtain different shapes and sizes. The multi-scale characteristics of the striation construct the dependence between different channels and make the network pay more attention to the image information characteristics of rain streaks and high-frequency regions. Finally, the feature reconstruction is performed through the convolutional layer to obtain a clear image of the removal of rain patterns. The experimental results show that the algorithm achieves a peak signal-to-noise ratio of 28.91, 36.86 dB and 35.14 dB, respectively on Rain100H, Rain100L and Rain12 open test sets, and a structural similarity of 95.0%, 99.0% and 97.1%. Experiments show that the objective evaluation indexes of the proposed algorithm are better than other comparison algorithms, and the subjective visual effect can be effectively improved, which can remove different density rain patterns while better preserving the details of the image.s of the proposed algorithm are better than other comparison algorithms, and the subjective visual effect can be effectively improved, which can remove different density rain patterns while better preserving the details of the image.

     

/

返回文章
返回