Research On Image Dehazing And Haze Image Quality Assessment

Hazy weather will not only do harm to people's health but also will affect traffic safety,because a city's surveillance system will be greatly weakened by the poor visibility.Therefore,researches on haze removal algorithms are quite necessary,since it is a huge matter of national security.Meanwhile,evaluating haze image quality subjectively and objectively help to improve the performance of post-processing algorithms on haze images,and will provide accurate and reliable comparison benchmark for these methods.In this paper,the research status and background about haze removal algorithms and hazy image quality assessment methods are firstly introduced,including atmospheric scattering model,dehazing algorithm based on prior information,dehazing algorithm based on machine learning and deep learning,the construction of haze image database,and the development of haze image quality evaluation methods.On this basis,we propose two improved haze removal algorithms and one novel haze image quality assessment method.The specific works are as follow:(1)A haze removal algorithm based on SLIC(Simple linear iterative clustering)super-pixel segmentation and cost function is proposed.This method is the improved version of a dehazing algorithm based on cost function,but with the extra design of super-pixel segmentation operation.Firstly,the hazy image is segmented using the SLIC method to acquire identical scene depth in the super-pixel set.Secondly,a hybrid cost function based on contrast and information entropy is adopted to calculate the optimal value of transmittance in the set by minimizing this function.At last,a quadtree partitioning method is used to compute the global atmosphere light value,then the dehazed image.(2)An improved dehazing algorithm called MSDehaze is proposed.This new method combines the merits of traditional MSCNN and AOD-Net methods that are both based on deep learning.MSDehaze net adopts a new single value objective function and uses SSIM based cost function,which has better coherence with subjective image quality assessment result.Meanwhile,the network structure is optimized to better evaluate network parameters.In general,the new method has better haze removal performance comparing to traditional neural networks.(3)A novel haze image quality assessment method based on sky recognition and dark channel prior is proposed.A new evaluation index is designed by solving the total attenuation rate of incident light under atmospheric scattering model,and by shielding the interference of sky region.Experimental results had shown the effectiveness of the improved method to accurately estimate the level of the hazy degree in images.This research aims to solve haze removal problem and hazy image quality assessment problem.Two improved dehazing algorithms and one novel haze image quality evaluation method are proposed.The performances of the new haze removal methods have been increased comparing to the traditional ones,and the newly designed haze image quality index reflects the subjective evaluation results of hazy images fairly well.The research in this paper also could provide some valuable experience and reference for future research.