Research On Image Based Visibility Inversion

Atmospheric visibility is a physical quantity closely related to human life.The phenomenon of low atmospheric visibility not only seriously affects people's travel safety,but also reflects the degree of air pollution to a certain extent,which has an indirect effect on the health of people.Traditional visibility measurement methods have some limitations,such as poor observability,high price of equipment,and cannot realize automatic monitoring.Also,the method of visibility measurement based on image has the difficulty of correcting the error of the formula parameters.The purpose of this paper is to find a better way to achieve the inversion of image visibility,so as to solve the above problems.The breakthrough of sparse low rank representation in the field of computer vision makes it an effective way of data representation.Convolution neural network performs well in the field of image processing.Therefore,based on these two methods,we have carried out the research of image visibility inversion.The main research work of this paper includes the following three parts:(1)Taking min-max normalization operation on the visibility value,the target area of the visibility image is retrieved,the scale invariant feature(SIFT)algorithm is used to register the image,and the registered images are stored in the database.(2)In this paper,sparse low rank expression is applied to image visibility inversion.After learning the sparse and low rank representation principle of the signal,a complete dictionary is constructed based on the eigenvector set of the training sample,and a new representation of the image features is obtained by learning,and the progress estimate is obtained by the regression method.The influence of different factors on the accuracy of visibility measurement is discussed,and the best plan is confirmed.Finally,we compare the sparse low rank representation with several other typical models.The experimental results show that the proposed method is superior to the other three models in terms of measuring time and accuracy.(3)Considering that the method of visibility inversion based on sparse and low rank representation needs to extract the feature of image in advance,it has a large amount of work,the method of visibility inversion based on convolution neural network is proposed.In this paper,a multi-layer network model is built on the basis of the classical Lenet network.Using the image directly as the input of the neural network,the corresponding visibility value is used as the label to train the model.The output layer is reformed to solve the regression problem,and the automatic visibility inversion without the manual extraction of image features is realized.We optimize the convolution layer number,the number and size of convolution kernel,then get the optimal model.Based on the measurement results of the CJY-1G visibility meter at the experimental base,compare the predictive value of the method of convolution neural network and of black body pixels.The experimental results show that the CNN method is superior to the DPA method,the visibility is best when the visibility is<2000m,The correlation coefficients between the measured values of the two methods and of CJY-1G are 0.9385 and 0.8957 respectively.Finally,analyze and summarize the proposed sparse low rank representation method and convolution neural network method,and take the convolution neural network method as the focus of the future research.