基于深度邻近连接网络的单幅图像去雨方法

雨天环境下的雨线导致图像内容被遮挡,严重影响人眼的视觉效果和后续系统的处理性能。目前主流的深度学习方法为了提升处理性能,均以复杂的网络结构和较大的参数量为代价,导致相关方法难以服务于实际应用。为此,文中提出一种新的深度邻近连接网络结构。它通过关注深度网络中所学特征图之间的关系,采用融合操作将邻近特征图进行连接,以获得更加丰富和有效的特征表示。实验数据表明,所提方法在3个公开合成数据集及真实有雨图像上的主客观处理效果、模型参数量和运行时间等相关性能都有所提升。在合成数据集Rain100H上的平均结构相似性(SSIM)值达到0.84,在合成数据集Rain100L和Rain1 200上的平均SSIM值分别达到0.96和0.91。在真实有雨图像上,所提方法在有效去除前景雨线的同时,能够保护更完整的背景图像信息,从而获得更好的主观视觉效果。相比于同时期的深度学习方法JORDER,文中方法在保证相近的处理效果的前提下,模型参数量和CPU运行时间分别降低了一个和两个数量级。实验数据充分说明,通过将网络中邻近特征图进行融合,能够获取更加有效的特征表示。因此,所提方法虽然仅使用较少的模型参数和简洁的神经...

Single Image De-raining Method Based on Deep Adjacently Connected Networks

Rain streaks result in the occlusion of image content,which seriously affects the human visual effect and the processing performance of subsequent systems.Existing deep learning-based methods improve de-raining performance at the expense of complex network structure and parameter burden,which makes these methods difficult for serving practical applications.Therefore,a deep adjacently connected network structure was proposed in this paper.By focusing on the relationship between learned feature maps in depth networks,a fusion operation is designed to connect the adjacent features to obtain rich and more effective feature representation.Experiments on three public synthetic datasets and real-world rainy images show that the proposed method improves de-raining performance on both subjective and objective evaluations.The average structural similarity(SSIM)value on the synthetic dataset Rain100H is 0.84,and the average SSIM values on the synthetic dataset Rain100L and Rain1200 are 0.96 and 0.91.In real-world rainy images,the proposed method can effectively remove the foreground rain streaks while protecting background image information to obtain better visual quality.Compared with JORDER,the proposed method achieves comparable de-raining results,and can reduce the model parameters and CPU runtime by one and two orders of magnitude,respectively.Experimental data demonstrate that fusing adjacent features in the deep network can generate more effective representation.Therefore,although the proposed method contains relative few parameters and simple neural network structure,it can still achieve better ima-ge de-raining performance and solve the problems of parameter burden and complex network structure in existing methods.Mo-reover,the network structure design scheme in this paper can also provide reference values for relative image restoration tasks based on deep learning.