Ordinal Relation and Relative Learning for Foggy Image Visibility Detection

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.