Application Research Of Convolutional Neural Network In Image Dehazing

Nowadays,hazy weather is becoming increasingly frequent and serious,thus affecting the quality of imaging devices by making photos less visible and unclear with haze on them.So the technology of image dehazing is becoming more and more important.With many relevant studies of image dehazing,a large number of methods have been proposed.However,every existing method has its own limitation.This thesis tries to use CNN(Convolutional Neural Network),which has been widely used and proved effective in the field of image processing,to study image dehazing and to design a new CNN that can work well for image dehazing.As there are not enough relevant studies about CNN on the problem of image dehazing at present,this thesis is only an exploration of the field with the hope of providing some reference for later studies.At first,this thesis makes an introduction of the physical model that describes the formation of hazing images.The model is the basis of image dehazing and it is what we base our study on.Meanwhile,this thesis also explains neural network,BP algorithm,and then introduces CNN and different layers of CNN further.These details are the basis to the study of this thesis.Next,after learning from some classic CNN,this thesis designs a CNN which is used to estimate the corresponding transmission map of hazing images for the problem of image dehazing.On the basis of the brief explanation of the design principle,the thesis explains the structure of CNN in detail and finally makes a simple theoretical analysis.Then,this thesis experiments on the designed neural network for image dehazing by using Caffe.After that,this thesis introduces its sample synthesis and the experiment process of the network.This thesis also tests the CNN for image dehazing,whose result is rather satisfying.In order to get better results,this thesis uses the widely adopted guided image filtering to smooth the transmission map of the corresponding hazing image that is the output of the CNN.At last,this thesis compares our method with two classic dehazing methods visually and quantitatively.In the process of quantitative comparison,by using the image datasets,which carries the information of scene depth,this thesis builds up a dehazing dataset.Not only can the dataset be used to quantitatively evaluate the result of the method of image dehazing,we can also use it as a training data to train the CNN for dehazing.The dehazing data is used by us to compare our method with two classic methods with the SSIM index.This thesis mainly makes a research on image dehazing by using CNN,and proposes an effective CNN for image dehazing.The designed CNN is used for image dehazing in real life.According to the test result,it works well both visually and quantitatively.Meanwhile,this thesis builds up a dehazing dataset by making use of image datasets with scene depth information.This dehazing dataset can be both used to quantitatively evaluate different dehazing methods and to train CNN for dehazing.