生成对抗映射网络下的图像多层感知去雾算法
发布时间：2019-04-09 06:57
【摘要】：雾霾常会影响获取图像的质量,单幅图像去雾是一个具有挑战性的不适定问题.针对传统的去雾方法存在去雾结果颜色失真、适用范围局限等问题,提出一种基于深度网络的去雾算法——生成对抗映射网络的多层感知去雾算法.在训练阶段中,利用生成对抗映射网络里判别网络与生成网络间对抗式训练机制,保证生成网络中参数的最优解;在测试还原过程中,先提取有雾图像中雾气相关特征,并利用训练得到的生成网络对提取特征进行多层感知映射,进而得到反映雾气深度信息的透视率,最终运用得到的透视率实现了图像去雾.实验结果表明,与同类算法相比,该算法能较好地还原出场景中目标的真实色彩,并抑制部分噪声,去雾效果明显.
[Abstract]:Haze often affects the quality of image acquisition, and the single image de-fogging is a challenging ill-posed problem. In order to solve the problems of color distortion of de-fogging results and limitation of application range in traditional de-fogging methods, an algorithm based on depth network for de-fogging is proposed, which is called multi-layer perceptual de-fogging algorithm based on anti-mapping network. In the training stage, the antagonistic training mechanism between the discriminant network and the generated network is used to ensure the optimal solution of the parameters in the generated network. In the process of test and restore, firstly, the fog-related features in fog images are extracted, and the multi-layer perceptual mapping of the extracted features is carried out by using the trained generation network, and then the perspective of fog depth information is obtained. Finally, using the perspective of the image to achieve de-fogging. The experimental results show that compared with the same algorithm, the proposed algorithm can restore the real color of the target in the scene, and suppress some noise, and the fog removal effect is obvious.
【作者单位】： 兰州理工大学电气工程与信息工程学院;西安交通大学电子与信息工程学院;
【基金】：国家自然科学基金(61365003) 甘肃省基础研究创新群体项目(1506RJIA031) 中国博士后科学基金(2014M550494)
【分类号】：TP391.41
[Abstract]:Haze often affects the quality of image acquisition, and the single image de-fogging is a challenging ill-posed problem. In order to solve the problems of color distortion of de-fogging results and limitation of application range in traditional de-fogging methods, an algorithm based on depth network for de-fogging is proposed, which is called multi-layer perceptual de-fogging algorithm based on anti-mapping network. In the training stage, the antagonistic training mechanism between the discriminant network and the generated network is used to ensure the optimal solution of the parameters in the generated network. In the process of test and restore, firstly, the fog-related features in fog images are extracted, and the multi-layer perceptual mapping of the extracted features is carried out by using the trained generation network, and then the perspective of fog depth information is obtained. Finally, using the perspective of the image to achieve de-fogging. The experimental results show that compared with the same algorithm, the proposed algorithm can restore the real color of the target in the scene, and suppress some noise, and the fog removal effect is obvious.
【作者单位】： 兰州理工大学电气工程与信息工程学院;西安交通大学电子与信息工程学院;
【基金】：国家自然科学基金(61365003) 甘肃省基础研究创新群体项目(1506RJIA031) 中国博士后科学基金(2014M550494)
【分类号】：TP391.41
【相似文献】
相关期刊论文 前10条
1 梁志刚,李坤成;基于个人微机的医学图像的采集与传输[J];中国医学影像技术;2003年10期
2 郭t，
本文编号：2454946
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2454946.html
