摘要：

在下雨天气中, 玻璃上的雨滴会对图像质量产生严重影响, 且目前的去雨滴方法过度依赖成对图像, 使得无监督图像雨滴去除面临较大挑战。针对这一问题, 提出一种域变换图像去雨滴方法。构建域变换网络(DTN), 通过有雨与无雨域之间的变换, 以无监督的方式实现图像的雨滴去除。同时, 通过引入快速傅里叶卷积(FFC)来设计生成网络和判别网络, 实现全局与局部特征的信息交互。在FFC中, 通过频谱变换(ST)对空间域和频域进行转换, 克服传统卷积神经网络(CNN)感受野不足的问题, 从而更好地感知细小的雨滴。在2个真实的雨滴测试集上进行去雨滴实验, 结果表明, 该方法在定量结果和视觉效果上均优于现有的先进方法。与改进前的U-Net+马尔可夫判别网络相比, 改进后的该方法在峰值信噪比(PSNR)和结构相似性指数(SSIM)上分别提升3.37 dB和0.031 3, 并且其能在去除雨滴的同时还原更多的图像纹理细节。

关键词: 图像雨滴去除, 快速傅里叶卷积, 无监督学习, 域变换, 频谱变换

Abstract:

Owing to the significant impact of raindrops on the quality of images under rainy conditions and the current over-reliance on paired images in existing rain removal methods, achieving unsupervised image rain removal remains a challenging research problem. To address this issue, this study proposes a domain transform image raindrop removal method. Building a Domain Transform Network (DTN) that transforms between rainy and rain-free domains to achieve unsupervised raindrop removal. Additionally, Fast Fourier Convolution (FFC) is introduced to design the generation and discrimination networks, enabling the interactions of global and local features. Within FFC, Spectral Transformation (ST) is employed to transform between spatial and frequency domains, overcoming the limited receptive field problem of a traditional Convolutional Neural Network (CNN) and enhancing the perception of subtle raindrops. Deraining experiments conducted on two real raindrop test sets demonstrate that our method outperforms existing advanced methods in terms of quantitative results and visual effects. Compared with the original U-Net plus Markov discriminant network, the improved method improves the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) by 3.37 dB and 0.031 3, respectively, and can restore more image texture details while removing raindrops.

Key words: image raindrop removal, Fast Fourier Convolution(FFC), unsupervised learning, domain transformation, Spectral Transformation(ST)