基于轻量化深度神经网络(LDNet)的图像去雾方法

Image Dehazing Method Based on Lightweight Deep Neural Network(LDNet)

在低光照雾霾场景下,图像质量严重下降。现有的深度学习去雾方法缺乏对低光照去雾后图像色偏的有效校正,且大多运行时间长且模型参数量大,在实际应用中不便部署。针对上述问题,以编解码网络结构为基础,提出了一种端到端、轻量化深度神经网络(LDNet)用于低光照雾霾图像去雾。该网络采用多尺度架构来获取不同层级的图像信息,以充分利用图像的深浅层特征;在此基础上,设计了轻量化多级特征融合模块和轻量化通道注意力模块提取各层级的特征信息,以解决常规模型在低模型参数量和低计算复杂度情况下特征提取能力差的问题;最后,联合均方误差内容损失和CIEDE2000色偏损失共同优化网络,进一步提高了轻量化网络的学习能力。试验结果表明,与现有的去雾网络相比,LDNet能有效恢复低光照雾霾场景下的有雾图像质量,且具有资源占用少、参数量小和运算量低的优点。

In low-light haze scenarios,image quality is severely degraded.In existing deep learning-based dehazing methods,it is lack of effective correction of image color bias after lowlight dehazing,and most methods need long run time and a large amount of model parameters.So they are inconve-nient to deploy in practice.Aimed at the above problems,an end-to-end lightweight deep neural net-work(LDNet)is proposed based on the encoder-decoder network structure to dehazing for low-light hazed images.In the LDNet,a multi-scale architecture is adopted to obtain different levels of image in-formation.Thus,it can make full use of the deep and shallow features of an image.On this basis,a lightweight multilevel feature fusion module and a lightweight channel attention module are designed to extract feature information of each level,in order to solve the problem about poor feature extraction ability of conventional models under the condition of low number of model parameters and low compu-tational complexity.Finally,the network is optimized by combining mean square error(MSE)loss with the color deviations loss of CIEDE2000.Thus,the learning ability of the lightweight network is further improved.Experimental result shows that compared with the existing dehazing network,LD-Net can effectively restore the quality of hazed images in low-light haze scenarios,and has advantages of low resource consumption,small quantity of parameters and low computation.