Research On Image Dehazing Algorithm Based On Dark Channel Prior

Today,with the rapid development of science and technology,computer technology has entered all walks of life at a rapid speed,the Internet of Things technology has developed rapidly,and the networking of monitoring equipment such as cameras has been widely used.At present,there are still hazy weather conditions in many cities.At this time,the pictures captured by the camera are always hazy,which brings inconvenience to many subsequent tasks,such as license plate recognition,pedestrian facial recognition,automatic driving road condition judgment,etc.small challenge.In this case,it is necessary to use image processing technology to process the picture,remove the fog layer on the image,and dig out the details hidden by the haze.Image dehazing has been the focus of research in the field of image processing in recent years,and many methods have been proposed one after another.These methods can be roughly divided into two categories: one is the traditional image dehazing methods,and This class of methods includes methods relying on atmospheric scattering models,image enhancementbased dehazing methods,and image prior-based dehazing methods.These three methods have their own unique parts and parts that are interlaced with each other;The second category is the dehazing method based on deep learning,which mainly realizes image dehazing by learning the mapping relationship between the haze map and the clear image.Most of the traditional methods are suitable for some specific images,while the deep learning method is applicable to a wider range,and the dehazing results are not prone to distortion,noise,etc.This thesis mainly proposes two different image dehazing methods based on dark channel prior in the traditional field and deep learning field,and achieved good results.The main research work is as follows:1.Aiming at the problem that the traditional methods are not completely dehazing and the noise in the image is amplified,this thesis proposes a single image dehazing algorithm based on an improved atmospheric scattering model.First,the traditional atmospheric scattering model is improved by adding a noise term to the model;then,it is improved on the basis of the dark channel prior method to calculate the transmittance,and the dense fog weight of the fog is added to solve the problem of underestimation of the transmittance;then,A new objective function is constructed,and the final haze-free graph is obtained by splitting the function and iterative optimization.A large number of experimental results prove that this method can suppress the amplification of noise while achieving effective dehazing.2.Most of the existing deep learning-based image dehazing methods directly learn the mapping relationship between the haze image and the corresponding haze-free image,without considering the characteristics of the haze image itself.Based on the above problems,this thesis proposes a Dark Channel Prior-Guided Image Dehazing Network(DCPDNet).First of all,in order to enhance the correlation of feature space information,a feature enhancement module(FEB)is constructed,in which two downsampling operations are set to reduce the computational complexity of the model,and the shallow features and deep features are combined by multiplication,to achieve the enhancement of image features.Then,in order to make the network focus more on the foggy area in the image,this thesis also proposes a guide map-based feature correction module(FCB),which uses the dark channel prior theory to construct the guide map,and evaluates the image transmittance through the guide map.The transmittance is used to adaptively amplify the eigenvalues of the hazy region,which can guide the network attention to the hazy region of the image.Finally,the shallow features are supplemented from the haze map using the residual structure,and the final output of the network is obtained through convolutional refinement.A large number of experiments prove that the method in this thesis can ensure the lightness and running speed of the model while achieving effective dehazing.