Image Dehazing By Feature Learning

With the rapid development of the national economy,the natural ecological environment has been destroyed to a certain extent.A large number of automobile exhaust and industrial exhaust emissions have a serious impact on air quality.This leads to the fact that more and more cities are affected by the haze disaster.In the fog,haze and other inclement weather,the atmosphere has more drops or dust particles,resulting in the light is scattered and absorbed in the transmission process.In this situation,outdoor visibility dropped significantly.Images taken under this condition are often blur and with low contrast.Moreover,the performance of the systems which depend on the clarity of the outdoor images cannot work normally.Under this background,it would be an effective solution to implement image haze removal as it has extremely important practical significance.At present,the mainstream algorithms mainly use a variety of haze-related color features to estimate the transmission map of the hazy image,and finally restore the haze-free image achieving the goal of single image dehazing.However,different color priors have their own limitations.In particular,they often fail to deal with the images with a large sky region or white objects.In order to solve the above problems,this paper proposes a single image dehazing algorithm based on feature learning,making it more suitable for more scenarios.The main research works and innovations are listed as follows:(1)By collecting a large number of outdoor haze-free images and the corresponding depth maps,we construct a standard dataset with the atmospheric scattering model for accessing the quality of image dehazing objectively.The dataset includes a large amount of hazy images and the corresponding haze-free images.We use different dehazing algorithms to remove the hazes from the hazy images,then compare the results with the ground truth haze-free images.In this way,we can acess and rank the performance of the dehazing algorithms.(2)We conclude most of single image dehazing algorithms which are based on physical model,then analyze the color priors or features used in the algorithms.Furthermore,we evaluate the effectiveness of the color features with the constructed ground truth dataset.(3)With the constructed dataset,we train and learn the haze-related structure and texture features by deep learning method.After that,we are able to combine these features and the traditional color features to achieve the goal of image dehazing.This method can deal with the hazy images which have white objects or sky regions successfully,improving the clarity of the image greatly.