Learning-based Image Dehazing Algorithms And Dehazed Image Assessment

As a common weather phenomenon,haze will reduce the visibility of the images captured by the camera,which will significantly lower its application value.Therefore,the task of image dehazing has important practicability.This paper focuses on image dehazing.Firstly,the advantages and disadvantages of the current image dehazing algorithms has been analyzed in this paper.Then,two different learning-based image dehazing algorithms and a pixel-level dehazed image quality assessment algorithm are proposed to contribute this field.The main research contents and achievements of this paper are as follows:First,in order to restore the hazy images to the clear images,this paper propose an end-to-end deep learning model for image dehazing.The proposed learning-based network does not require estimation of the transmittance and the atmospheric light value of the hazy image,but the residual between the clear image and hazy image is directly estimated,and then the hazy image is used to add the residual map to get a clear image.Second,inspired by the traditional image hazing algorithms,this paper proposes a new activation function called the Reverse Parametric Rectified Linear Unit(RPRe LU)to improve the dehazing performance of the network.Third,in order to solve the problem that the traditional learning-based image dehazing algorithms need paired images to train,this paper propose an image dehazing model,named Cycle CAGAN,based on the Generative Adversarial Network,which does not need pair data to train.The proposed Cycle CAGAN can not only be used to image dehazing task,but also can be used to simulate hazy images from clear images.Fourth,this paper proposes a new loss function based on the prior knowledge.It can help to improve the defogging effect of the image defogging model based on the generative countermeasure network.Fifth,in order to evaluate the performance of the image dehazing algorithm quantitatively,this paper proposes a pixel-level dehazed image quality assessment algorithm,which can not only evaluate the whole image objectively and fairly,but also evaluate a specific pixel in the image.Sixth,this paper proposes a new depth estimation algorithm which can roughly estimate the depth of the hazy image combined with the characteristics of the hazy image.The depth of the image estimated by the algorithm can help the proposed image quality assessment algorithm to evaluate the performance of the dehazed image more accurately.