PD-Net：基于金字塔池化模块改进的Dehaze-Net

PD-Net: Improved Dehaze-Net Based on Pyramid Pooling Module

摘要: 对单幅雾天图像进行去雾是一个困难的任务，本文提出了一种名叫PD-Net的神经网络模型用于单幅图像去雾。该模型在Dehaze-Net模型中引入了金字塔池化模块且额外添加残差块。金字塔池化模块增强了模型对全局信息的提取，残差块有效地抑制了梯度消失现象。在仿真实验中，本文使用RESIDE数据集中的室内合成有雾图片作为实验数据集，定量分析比较了各种去雾方法对实验数据集的去雾结果，PD-Net模型展现了良好的性能。在真实实验中，对RESIDE数据集中的室外真实有雾图片进行去雾分析。真实实验的结果表明，相较于其他算法，PD-Net模型在大面积的天空区域及图片的细节处有更好的效果。

Abstract: It is a difficult task to remove haze from a single input image. In this paper, we present a neural network model called PD-Net for dehazing single image. In this model, Pyramid Pooling Module is introduced into Dehaze-Net model, and Residual blocks are added. The pyramid pooling module enhances the extraction of global information, and the residual block effectively suppresses the disappearance of gradient. In the simulation experiment, we use the indoor synthetic foggy pictures in the RESIDE set as the experimental set, quantitatively analyze and compare the defogging results of various defogging methods on the experimental set, and the PD-Net model shows good performance. In the real experiment, we defog and analyze the images which is the outdoor real foggy pictures in the RESIDE set. The results of real experiments show that compared with other algorithms, PD-Net model has better effect in large-area sky area and details of picture.