基于多流扩张残差稠密网络的图像去雨算法

针对传统图像去雨算法未考虑多尺度雨条纹及图像去雨后细节信息丢失的问题,提出一种基于多流扩张残差稠密网络的图像去雨算法,利用导向滤波器将图像分解为基础层和细节层。通过直接学习含雨图像细节层和无雨图像细节层的残差来训练网络,缩小映射范围。采用3条带有不同扩张因子的扩张卷积对细节层进行多尺度特征提取,获得更多上下文信息,提取复杂多向的雨线特征;同时,将扩张残差密集块作为网络的参数层,加强特征传播,扩大接受域。在合成图片和真实图片上的实验结果表明,所提算法能有效去除不同密度的雨条纹,并较好地恢复图像细节信息。通过对比其他算法,证明了所提算法在主观效果和客观指标上都有提升。

Image Deraining Algorithm via Multiflow Expansion Residual Dense Network

The traditional image rain removal algorithms do not consider multiscale rain streaks and often result in loss of detailed information after the image is derained.To solve these problems,an image rain removal algorithm based on a multiflow expansion residual dense network is proposed in this study.In this algorithm,a guided filter is used for decomposing an image into a base layer and a detail layer.The mapping range can be considerably reduced by training the network with the residuals present between the rain and rainless image detail layers.Three dilated convolutions with different expansion factors are used to perform multiscale feature extraction on the detail layer to obtain more context information and extract complex and multidirectional rainline features.Further,the expanded residual dense block,which is the parameter layer of the network,is applied to enhance the propagation of features and expand the acceptance domain.The experiments conducted on synthetic and real pictures show that the proposed algorithm can effectively remove rain streaks with different densities and restore the detailed information present in an image.When compared with other algorithms,the proposed algorithm is better in terms of subjective effects and objective indicators.