利用序数关系实现相对学习的雾图像能见度检测算法  

Ordinal Relation and Relative Learning for Foggy Image Visibility Detection

作　　者：吴琪 汪小武[1] 章军[1] 夏懿[1] 陈鹏 阎庆[1] Wu Qi;Wang Xiaowu;Zhang Jun;Xia Yi;Chen Peng;Yan Qing(School of Electrical Engineering and Automation,Anhui University,Hefei 230061;School of Computer Science and Technology,Anhui University,Hefei 230061)

机构地区：[1]安徽大学电气工程与自动化学院,合肥230061 [2]安徽大学计算机科学与技术学院,合肥230061

出　　处：《计算机辅助设计与图形学学报》2020年第12期1938-1947,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金青年科学基金(61602002).

摘　　要：针对雾图像的能见度检测在自动驾驶、气象预报等领域有巨大的研究和应用价值,而现有算法忽略了图像之间的序数关系的问题,提出一种利用序数关系进行相对学习的能见度检测算法.首先利用数据集中隐藏的序数关系约束网络学习过程,并将这种约束简化为图像特征三元组的相对距离关系,使得图像在特征空间的分布遵循序数规律;然后提取训练样本的能见度特征来构造查询库,通过在特征查询库中寻找近邻图像,得到测试样本的能见度检测结果.在公开的合成雾图像数据集(SF,FROSI)和构建的真实雾数据集(RDF)上进行了验证,实验结果表明,该算法取得了比现有的多分类、有序多分类等深度学习算法更好的效果,并且算法在训练过程中更加稳定,训练数据量要求更少,具有良好的稳健性和广阔的应用空间.The visibility detection of fog image has great value in the field of automatic driving,weather forecast and so on.However,existing methods ignored the relative information among images.Inspired by this,we propose a deep learning method based on the ordinal relation and triplet relative learning to conduct visibility detection of foggy image.The method uses the ordinal relation hidden in the data to constrain the network learning process and simplifies it to the relative metric of image triplet,which makes the distribution of the image in the feature space follow the law of ordinal rule.The visibility characteristics of the training samples are extracted to construct the query library,and the visibility labels of the test samples are obtained by finding the adjacent images in the feature query library.The method was validated on two public synthesized foggy image dataset(SF/FROSI)and our real dense foggy image dataset(RDF).Experimental results show that the proposed method has better effect than some baseline deep learning methods such as deep multi-classification or ordinal multi-classification.In addition,the method is more stable in the training process and requires less training data,which has excellent robustness and wide application space.

关 键 词：能见度和成像 机器视觉算法 深度学习 序数关系 相对学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]