Integrating Atmospheric Model And Deep Convolutional Network For Image Dehazing

The hazy weather affects the performance of outdoor imaging equipment.It cut down the quality of image acquired by outdoor imaging equipment,and influence the accuracy of image content analysis algorithms.Thus,image dehazing task has become a research hot spot in the field of image enhancement.Although the image haze removal algorithms have made some progress,this task is still challenging.At present,the image dehazing methods can be divided into three categories: methods based on contrast enhancement,prior knowledge-based traditional methods and neural network-based methods.Early dehazing methods based on contrast enhancement can solving simple hazy image.However,these methods are computationally intensive and inefficient.Prior knowledge-based methods have strict theoretical guarantees that require complex and accurate prior designs.The complex network models are needed by neural network-based methods.They rely heavily on large amounts of training data,and have poor generalization capabilities.A single image dehazing algorithm is proposed which integrating atmospheric model and deep convolutional network,aiming at solving the low contrast,low definition and high noise of hazy image.The excellent learning ability of deep convolutional network has been fully utilized in the proposed algorithm without complicated prior design and huge training data.Firstly,an optimization model is established based on transmission which is the crucial variable in atmospheric model.A plug-and-play optimization strategy is adopted to establish a solution which integrating deep convolutional neural network.The transmission can be smoothed to restore the sharp image from hazy image.Secondly,we also focus on the couple relationship between clear image and transmission,and establish the optimization model which joint two variables for enhancing haze image interfered by noise.For the purpose of significantly improving the visual quality of image,the deep neural network is introduced to guide the transmission smoothing and image denoising in optimization process,respectively.Extensive experiments have verified that the algorithm can effectively enhance the haze image.Finally,the proposed algorithm is generalized to other two low-level vision tasks(underwater image enhancement and low-light image enhancement).The strong generalization ability and robustness of the proposed method can be found compared with state-of-the-art methods.In general,a single image dehazing algorithm has been established for obviously improving the quality hazy image,which fully utilizes the learning ability of deep convolutional network according to atmospheric model.Then the proposed algorithm can generate better enhancement results when applied to underwater image and low-light image enhancement tasks.