光学学报, 2019, 39 (10): 1010001, 网络出版: 2019-10-09   

基于多尺度卷积神经网络的单幅图像去雾方法 下载: 2079次

Single Image Dehazing Method Based on Multi-Scale Convolution Neural Network
作者单位
兰州交通大学电子与信息工程学院, 甘肃 兰州 730070
摘要
针对传统的单幅图像去雾算法容易受到雾图先验知识制约及颜色失真等问题,提出了一种基于深度学习的多尺度卷积神经网络(CNN)单幅图像去雾方法,即通过学习雾天图像与大气透射率之间的映射关系实现图像去雾。根据大气散射模型形成雾图机理,设计了一个端到端的多尺度全CNN模型,通过卷积层运算提取有雾图像的浅层特征,利用多尺度卷积核并行提取得到有雾图像的深层特征,然后将浅层特征和深层特征进行跳跃连接融合,最后通过非线性回归得到雾图对应的透射率图特征,并根据大气散射模型恢复出无雾图像。采用雾图数据集对该模型进行训练测试。实验结果表明,所提方法在合成有雾图像和真实自然雾天图像的实验中均能取得良好的去雾效果,在主观评价和客观评价上均优于其他对比算法。
Abstract
Since the traditional single image dehazing algorithm is susceptible to the prior knowledge constraint of hazy image and color distortion, this paper proposes a multi-scale convolutional neural network (CNN) single image dehazing method based on deep learning, which realizes image dehazing by learning the mapping relationship between hazy image and atmospheric transmission. According to the hazy image forming mechanism of atmospheric scattering model, an end-to-end multi-scale full CNN model is designed. The shallow layer features of hazy image are extracted by convolution layer operation, and then the deep features are extracted by multi-scale convolution kernel in parallel. Then the shallow layer features and deep features are fused by jump connection. Finally, the non-linear regression method is used to obtain the corresponding transmission features of the hazy image. According to the atmospheric scattering model, the haze-free image is restored. The model is trained by using hazy image data sets. The experimental results show that the proposed method can achieve good dehazing effect in the experiments of synthesizing hazy images and real natural hazy images. The proposed method is superior to other contrast algorithms in subjective and objective evaluations.

陈永, 郭红光, 艾亚鹏. 基于多尺度卷积神经网络的单幅图像去雾方法[J]. 光学学报, 2019, 39(10): 1010001. Yong Chen, Hongguang Guo, Yapeng Ai. Single Image Dehazing Method Based on Multi-Scale Convolution Neural Network[J]. Acta Optica Sinica, 2019, 39(10): 1010001.

本文已被 11 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!