针对现有煤矿尘雾图像清晰化算法存在的过增强现象和适用性不足等问题,提出一种基于深度融合网络的清晰化复原算法。深度融合网络主要包括3个部分，即图像预处理模块、特征融合模块以及图像输出模块。图像预处理模块基于对比度增强函数、亮度增强函数和伽马校正函数对输入图像进行处理，获取表征不同增强方式及程度的图像序列。由于图像尘雾清晰化需要同时考虑图像的局部信息和全局信息,在空间金字塔池化和上下文信息聚合网络的基础上提出了能够实现双向的上下文信息提取的双金字塔模块，该模块包括2个空洞卷积的串联子块,其中1个子块是以对多个尺度的空洞卷积按尺度由小到大进行串联组成,另1个子块是以对多个尺度的空洞卷积按尺度由大小进行串联组成。图像输出模块主要对特征融合层获取的特征进行处理,从而输出三通道图像,即为最终的尘雾清晰化的图像。为了获取训练数据，本文在煤矿井下清晰图像的基础上基于尘雾图像形成机理构建了较大规模的训练数据集。在训练的过程中，采用了最小平方误差损失函数和基于VGG网络的内容损失函数对网络进行优化。为评价本文提出的基于深度融合网络的清晰化算法的有效性,选取其他6种有代表性的清晰化算法进行对比。实验结果显示，本文算法在主观评价和客观评价方面都优于上述算法，表明本文算法能够有效解决过增强现象,并提升煤矿图像的清晰度和可视化效果。

A clear restoration algorithm based on deep fusion network is proposed to solve the problem of over enhance- ment and insufficient applicability of the existing dust and fog image clear restoration algorithm. The deep fusion net- work mainly includes three parts,namely,the image pre-processing module,the feature fusion module,and the image output module. The image pre-processing module processes the input image based on the contrast enhancement func- tion,the brightness enhancement function,and the gamma correction function to obtain an image sequence that charac- terizes different enhancement modes and degrees. Because the local information and global information of image need to be taken into account,this paper proposes a double pyramid module which can realize a dual-path context informa- tion extraction on the basis of spatial Pyramid pooling and context information aggregation network. The module consists of two series sub-blocks of dilated convolution,one is composed of a series of small to large scale dilated convolu- tion on multiple scales,and the other is composed of a series of large and small scale void convolution on multiple scales. The image output module mainly processes the features acquired by the feature fusion layer,thereby outputting a three-channel image,that is,a clear image. In order to obtain the training data,this paper builds a large-scale train- ing data set based on the dust fog image formation mechanism with the clear coal mine images. In the process of train- ing,this paper uses the least square error loss function and the content loss function based on VGG network to optimize the network. In order to evaluate the effectiveness of the proposed algorithm based on deep fusion network,six other representative clearing algorithms are selected for comparison. The experimental results show that the proposed algo- rithm outperforms the other six algorithms in subjective evaluation and objective evaluation,which indicates that the proposed algorithm can effectively solve the over-enhancement phenomenon and improve the clarity and visualization of coal mine images.