Feature Learning Based Single Image Dehazing Algorithm

Haze is a natural phenomenon that occurs when water vapor near the ground condenses.Due to the presence of haze,the outdoor monitoring system and other applications cannot be carried out normally,and the target can not be identified,extracted and processed.In the process of identification,important details may be omitted,which is not conducive to the further development of the research work.Therefore,the research of dehazing technology is very important and meaningful.Image dehazing technology can be generally divided into three categories: dehazing algorithm based on image enhancement,dehazing algorithm based on image restoration,and dehazing algorithm combining image enhancement with image restoration.The early dehazing technology is mainly based on image enhancement.This dehazing method is relatively mature,and it achieves the purpose of dehazing by improving the visual effect,but it is easy to introduce noise and produce distortion.In view of this situation,the researchers combined with the physical model of atmospheric scattering and proposed a dehazing algorithm based on image restoration.The dehazing image recovered by this algorithm is more natural and the details are well preserved,but it is easy to introduce noise in the local area,such as color deviation.Considering the advantages and disadvantages of dehazing algorithm based on image enhancement and dehazing algorithm based on image restoration,the two algorithms are combined to remove haze.In this paper,two different kinds of dehazing algorithms are proposed.The first algorithm combines image enhancement technology with image restoration technology to remove haze,and the second algorithm is based on image restoration technology to remove haze.The main contributions of this article are as follows.1.To solve the problem that the existing dehazing algorithm cannot be applied to all hazy scenes,a single image dehazing algorithm based on HSV model and Retinex theory was proposed.Firstly,the haze map training data set is constructed,which includes as many haze scenarios as possible.Next,the image is transferred from the RGB space to the HSV space.According to the Retinex theory,the brightness channel map is transformed with SSR,and the features of each haze map are extracted at multiple scales,adopting different scales can effectively avoid the omission of important information.In the RGB space,the largest channel feature map of the original haze map is taken,and the anti-color processing and gaussian filtering are carried out on it.Then,the extracted features and corresponding tags are sent to Support Vector Machines(SVM)to train the regression model.Finally,for any hazing image,features are extracted,transmission is calculated and optimized by the regression model learned above.According to the optimized transmission and atmospheric light estimation,the haze-free image can be restored.Experimental results show that the method is effective in removing haze.2.In order to solve the problems of incomplete dehazing,color distortion or oversaturation in the existing dehazing algorithms,a single image dehazing algorithm based on multi-scale filtering and local mode is adopted.Firstly,for a single hazy image,multi-scale filtering and edge detection were carried out in R,G and B channels respectively,meanwhile,dark channel features of the original hazy image were extracted,and features and corresponding labels were sent to SVM for training to obtain regression model 1.Next,the Uniform Pattern LBP features and corresponding labels of the original haze image were sent to SVM for training to obtain the regression model 2.Then,the two regression models are used to calculate two different transmission of any haze diagram.Finally,the two transmission are fused proportionally to get the best transmission.This algorithm can effectively suppress color bias problem.