Single Image Haze Removal Based on Mean Channel Prior _MCP_ and U_NET Based Encoderdecoder Architecture

Abstract

Haze is an extreme weather condition arises due to natural phenomenon and newlinehuman activity, in which dust, aerosols and fine particles suspended in air can newlineseverely hamper the visibility of the objects. Several applications include object newlinedetection, security surveillance and photography get affected due to unavoidable hazy newlineweather condition. So removing haze can leverage several applications to progress newlinefurther. newlineRemoving haze is termed as dehazing and it is extremely difficult due to not newlinehaving well defined mathematical model. Two parameters namely transmission newlinecoefficient and airlight estimation is essential for reconstruction of haze free image. In newlinethis work, we propose, computationally efficient statistical observation called Mean newlineChannel Prior (MCP) for estimating transmission coefficient, based on the newlineexperimental evidence that the concentration of haze is proportionally related to newlineaverage value of the three channels of the color image namely red, green and blue. newlineFast guided filter is used to refine the transmission coefficient without losing the edge newlineinformation. Airlight is estimated based on top one percent brightest pixels of the newlineimage. The experimental results have shown that the simple MCP method along with newlinefast guided filter have shown impressive results. newlineRecently deep learning models were performing better, importantly for the problems newlinewhich cannot be explicitly defined mathematically. In this work, we make use of deep newlinelearning architecture for estimating transmission map (or) haze depth map and fast newlineguided filter Airlight estimation, noticed performance improvement over the previous newlinemethods. Further, to improve the performance of the model,we propose an end-toend newlineintegrated encoder-decoder based deep learning architecture for both the newlineparameter estimation (transmission coefficient and Airlight) and reconstruction of haze free image.Hazy image classifier is also modeled for automatic detection of newlinehazy images newline