Research And Implementation On Instant Haze Removal From A Single Image

In recent years,due to the environmental pollution,bad weathers happens frequently.Especially the haze weather is increasing.Bad weathers have much effect on imaging,which seriously affected the work of outdoor vision systems such as outdoor exploration system,outdoor surveillance system,automatic navigation system,the validity and reliability of traffic management system.In all kinds of weather conditions,fog is one of the biggest influence on the image.Due to the atmosphere scattering,images under the foggy weather conditions are degraded,appearing low contrast and color faded.It brings trouble to the traffic safety and outdoor visual observation.Therefore,the research of dehazing image technology has important practical significance.This paper mainly aims for image improvement for outdoor vision system under bad weather condition.For the purpose of solving the real time engineering problem,after study of the atmospheric scattering theory and image contrast degradation mechanism under the foggy weather,and the existing haze removal technology,we propose several new methods focusing on the instant single image removal based on imaging model and non-model.This paper detailed research content as follows:Local histogram equalization can usually effectively increase image contrast,but for the big picture in the real-time video the calculation is a great cost.And it has block effect.To solve these problems,this paper puts forward the self-adaptive partially overlapping histogram equalization method.It adaptively select the move length and the window size of histogram equalization based on image feature.It realize the double objectives of contrast enhancement and calculating reduction(through reducing the overlap part).At the same time with themethod of weighted average according to the number of overlapping it greatly reduced the block effect.Totally it realize the fast haze removal from a single image.Retinex is based on color constancy theory.It can enhance image effectively especially for dark images.But for haze images(generally bright)it has deficiency.It probably cause the ‘halo' phenomenon and color distortion.Considering haze will not impact the whole gradient distribution in images,this paper proposed a TV-Retinex combined with the gradient leading.The gradients of image edge are enhanced to improve image's details and eliminate the phenomenon of the halo at the same time.The algorithm is originally desiged to need a clear image as a gradient reference.In some condition it cannot provide clear images,we put forward the improvement edition using the single haze image through the iterative process.Haze image restoration method based on the atmospheric scattering imaging model have common disadvantages of computing complexity or needing additional conditions.Considering the fine estimating of the transmission and its amount of calculation,on the one hand guide filter is introduced to keep the image edge by using original image,it greatly reduce the complexity of the algorithm and reduce the operation time;On the other hand,we use tolerance parameter to improve calculation tolerance degree,avoided the distortion caused by calculation.This article also proposed to improve the airlight estimation method.It achieve rapid result from a single haze image and in a certain extent,to broaden the scope of the dark grey prior to fog method.According the dichotomy mechanism of atmospheric scattering imaging model,through the analysis of different degrees of the two parts in the model under different weather,we puts forward a simplified transmission estimation method based on scene depth distribution.Design high-pass filter of fog image pretreatment based on the fog feature of low frequency which prevented the usual dark results by general image restoration methods.Finally on the basis ofthe imaging model for single haze image it fast recovered fast normal visual images without haze.The results of the methods' experiments are analyzed and evaluated through contrast,the information entropy and the average gradient image evaluation etc.,based on the image dehazing effect and efficiency of the algorithm,the stability of subjective and objective evaluation.Finally the paper summarized the full text,and the further research and development direction were discussed from the real-time performance and practicability of the technology.