Research On Visibility Detection Based On Data-Driven

Visibility is an indicator of atmospheric transparency,and an important element of meteorological observation.It's important to detect visibility accurately and efficiently.Traditional visibility measurement methods have problems such as poor objectivity,limited frequency of observation,and high price.Therefore,this thesis studies how to detect the scene visibility with images,and adopts methods of visibility detection based on data-driven.The image features that reflect the change of visibility are extracted and the visibility detection model is trained through the visibility-annotated data to achieve the visibility detection.The main work of thesis are as follows:Firstly,addressing the problem of it's difficult to reflect the impact of visibility attenuation on imaging comprehensively with a single feature,the saturation and image interest region contrast feature are extracted,and the multiple regression model is used to train the mapping relationship between the image features and visibility,which reduces the complexity of model and obtains fine detection accuracy.Secondly,aiming at the problem that the manual setting feature usually cannot reflect the various effects of the visibility attenuation on the image thoroughly and it is difficult to get a large volume of visibility-annotated training samples,a method of visibility detection based on transferring an encoding network directly is proposed.Each image is divided into several subregions,which are encoded to extract feature with the pre-trained deep neural network.Then,a support vector regression models is utilized to map the features to the visibility after trained with the encoded feature.After that,fusion weight of each sub-region is evaluated according to the error analysis of the support vector.Experimental results show that the method reflects the effect of the visibility attenuation on the image and obtain better detection accuracy at the same time.Finally,to improve the representation power of image features,the encoding network is replaced based on the method of transferring an encoding network directly,and the pre-trained neural network is fine-tuned with the current detection results to optimize the parameters of the coding network.Compared to the method of transferring an encoding network directly,the method of this paper obtains better detection accuracy and meets the requirements of daily observation application.