A Single Image Dehazing Method Based On Sky Recognition

Images taken in bad weather are often subject to the atmospheric absorption and scattering,resulting in poor visual effects,such as reduced contrast,color fidelity,and blurred details.However,the majority of applications in computer vision are based on the assumption that the input images or videos are clear so that degraded images can cause great errors inevitably.Using these degraded images as input may cause catastrophic errors for the subsequent interpretation,analysis,recognition.Hence,researches on high-quality,rapid and universal dehazing algorithms are of theoretical significance and practical value.Recently,benefiting from the atmospheric scattering model proposed by Narasimhan and Nayar,varieties of single image dehazing methods have gradually matured.In this paper,we first analyzed the dark channel prior(DCP)dehazing algorithm,which is widely recognized in the field of single image dehazing.Dark channel prior is invaild for sky region,which leads to serious color shift and over-enhancement in the recovered image.In order to solve theses drawbacks,this paper proposes a single image dehazing method based on sky recognition.Compared with the existing techniques,the main contributions can be summarized as follow:Firstly,we put forward a novel sky region features prior,which is a statistic prior based on thousands of images.Based on the proposed prior,the sky region features map is constructed.Through the morphological opening operation,the sky region feature map is optimized to eliminate small noise and accurately identify the sky region in the image.Experimental results show that the proposed sky region features prior can stably identify the sky region in the image.Secondly,aiming at addressing the problem of global atmospheric light misjudgment existing in the state-of-art dehazing algorithms,an atmospheric light estimation method based on the sky region features is proposed.Considering that atmospheric light should exist in the sky region and have low saturation,the accuracy of atmospheric light estimation is improved.Experimental results show that the improved atmospheric light estimation method can locate the location of atmospheric light,which makes the estimation results more more precisely.Finally,the transmission values in sky region where the dark channel prior failed will be smaller than actual values,which introduces much noise and causes color distortion in the recovered sky region.We propose a new transmission estimation method based on average saturation prior,and estimate the transmission within sky region and non-sky region independently.The transmissionmodel is refined via the guided total variation model and the final clear image can be recovered through the atmospheric scattering model.Experimental results verify that the proposed transmission estimation method can reduce the over-enhancement of the sky region and improve the efficiency of dehazing,so that the algorithm can be applied to real-time processing.