基于神经网络的深浅层融合图像去雾算法  

Dehazing Algorithm of Deep and Shallow Fusion Images Based on Neural Network

作　　者：张玮韬 杜楷文 王永顺[1] ZHANG Wei-tao;DU Kai-wen;WANG Yong-shun(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《兰州交通大学学报》2022年第3期92-99,共8页Journal of Lanzhou Jiaotong University

摘　　要：由于受雾霾、沙尘等天气的影响,空气中存在的杂质颗粒对光线产生散射,导致室外和室内拍摄图像呈现出灰白化,阻碍了视觉系统智能化的发展.为了解决现有的深度学习图像去雾算法存在的去雾不彻底、透射率计算不精确的问题,提出了改进的卷积神经网络去雾算法.该网络先进行浅层特征提取,再经过多尺度卷积实现深层特征提取,最后再加入金字塔池化网络确保卷积网络输入图像尺寸一致,实现深浅层特征融合.改进后的网络能够获取特征图更多的信息并且防止丢失一些图像的细节信息,从而得到更精确的透射率图.仿真实验结果表明:改进后的算法可以提高图像去雾性能,有效地改善去雾图像出现的偏色、失真、去雾不彻底等问题,使去雾后的图像更加自然并呈现良好的视觉效果.Due to the influence of haze,sand and other weather phenomenon,the impurity particles in the air scatter the light,which causes the outdoor and indoor images to be grayed out and hinders the development of intelligent vision system.In order to solve the problems of incomplete dehazing and inaccurate transmittance calculation in deep learning image dehazing algorithms,an improved convolutional neural network dehazing algorithm is proposed.The network first performs shallow feature extraction,then performs deep feature extraction through multi-scale convolution,and finally a pyramid pooling network is added to ensure that the input image size of the convolutional network is consistent,and the fusion of deep and shallow features is realized.The improved network can obtain more information from the feature map and prevent the loss of some image details,thereby obtaining a more accurate transmittance map.The experimental results show that the improved algorithm can improve the image dehazing performance,in which the color cast,distortion and incomplete dehazing of the dehazing images have been effectively improved,so that to make the dehazed image more natural and present a good visual effect.

关 键 词：图像去雾 深度学习 神经网络 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]