多级特征逐步细化及边缘增强的图像去雾

Multi-level features progressive refinement and edge enhancement network for image dehazing

为设计合理有效的神经网络框架,提高去雾算法的精度,保留完整的边缘细节,提出了常微分方程(Ordinary Differential Equation,ODE)启发的多级特征逐步细化及边缘增强的去雾算法。利用多级特征提取子网络,从输入雾图中提取出包含细节信息的低级特征和包含语义信息的高级特征,用于后续去雾结果的逐步细化。受残差网络框架与ODE求解策略关联性启发,依据两步两阶的蛙跳方法Leapfrog设计出Leapfrog模块,并串联多个Leapfrog模块,模拟ODE离散的逼近求解过程,构造逐步细化的去雾子网络。此子网络中,每个Leapfrog模块在交替输入的低级/高级特征的互补信息引导下,不断细化前一个Leapfrog模块估计的去雾结果。受二阶微分算子实施边缘增强的启发,边缘增强子网络利用预训练的UNet估计最后一个Leapfrog模块的去雾图像边缘,并叠加到此去雾图像上得到增强边缘,保留细节的最终去雾结果。实验表明,在真实图像及合成图像上,本算法均能取得较好的去雾效果,且在视觉评价和客观评价方面优于已有的去雾算法,与EAAN相比去雾精度提高了5%,运行时间仅有0.032 s,能有效用于图像去雾的工程实践中。

An ordinary differential equation(ODE)-based multi-level feature progressive refinement and edge enhancement network is proposed for image dehazing to provide an effective convolutional neural network framework with an algorithm designed to preserve edges while improving accuracy.The study mainly comprises subnetworks of multi-level feature extraction,ODE-based progressive refinement,and edge enhancement.First,the multi-level features extraction subnetwork is leveraged to extract low-level features with detailed information and high-level features with semantic information from hazy images,to enable the subsequent dehazing operations.Second,a novel Leapfrog module is designed based on the relationship between the residual framework and ODE solver by cascading Leapfrog modules to model an approximation solution for ODEs.Finally,the progressive refinement subnetwork is developed.Notably,each Leapfrog module refines the output of the previous Leapfrog with alternative low/high-level features.Finally,motivated by the effectiveness of edge enhancement via second-order differential operators in the edge enhancement network,the edge of the dehazing result predicted by the last Leapfrog module is calculated using the pretrained UNet and added back into dehazing to enhance edges and preserve details.The experimental results demonstrate that the proposed method outperforms the existing methods on both synthetic images and real images quantitatively as well as qualitatively.The dehazing accuracy is improved by 5%and the runtime is only 0.032 s.Hence,the proposed method can be incorporated into practical dehazing applications in engineering.