Research Of Polarization-based Analysis Of Hazy Image

The analysis of images acquired in poor-visibility conditions is an importantbut challenging task in Computer Vision. Haze as well as fog are such problemsand manifest increasing attention for several years. In hazy condition, outdoorvisual systems are often suffered from problems of low contrast, imaging blurringand etc because of atmospheric scattering particles. Such problems will bring usdifficulties to collect useful information.In this paper, our main objective has been to discuss preliminarypolarization-based applications in the analysis of hazy image, by combining imageprocessing technique with polarization theory. Here, we mean the analysis by twoaspects: hazy image restoration and segmentation. Mainly, the thesis containscontents as follows.(1)Extract polarization information of objects. In hazy conditions, the measuredpolarization state is determined not only by the object’s reflection but also byambient illumination and cannot directly indicate the properties of the objectsurface. Thus, we propose a decorrelation-based method to separate the degree ofobjects with the degree of airlight, benefiting from a simple principle that the scenedepth and the object signals are statistically uncorrelated. Simultaneously, thetransmission map that indicates the depth information is inferred in the process.(2)Recovering clear-day scenes using estimated transmission map. Analyzingthe hazy degradation model, we present a suitable restoration formula and obtain agood dehazing result. Our method allows both the airlight and the object reflectionto be partially polarized. It thus outperforms and generalizes priorpolarization-based method, which had assumed the object reflection to beunpolarized.(3)Partition the hazy image into segments. Scene segmentation is mainlydependent on visual features extracted from images, such as intensity, color andtexture. However, the feature extraction from hazy image becomes difficult andunstable. To get more reliable and better result, a feature vector that suitable tohazy image segmentation is introduced and fed to a well-known graph-basedsegmentation algorithm. The vector is composed of color intensity, polarization character and depth information. After various aspects of comparative experiments,we conclude that the proposed feature vector is more effective and robust than theusual features for hazy scene segmentation.