Illumination Normalization Of Remote-sensing Images

Illumination condition is one of the most important conditions during imaging process, which affects image quality. Illumination homogeneity is required in many applications, such as MTF computation according to comparisons of high and low resolution images, remote-sensing image mosaics, face recognition and color consistency so as to many researchers do research on illumination homogeneity. Our destination is to propose a method to normalize illumination among multiple images. Thus the images that not satisfy illumination homogeneity can be used in many applications after illumination correction with our method.Our research approach is to correct the illumination in a single image by gradient domain method firstly and then normalize the illumination among multiple images by our proposed method. Finally we assess the illumination homogeneity by subjective vision and an objective quantitative assessment parameter to show that our method is effective. We find that there are several kinds of illumination conditions among remote-sensing images by observing large amounts of images:dark, bright or low contrast. So it is necessary to correct illumination in a single image before normalizing multiple images. We choose gradient domain method because the experimental results of this method are better than those of other methods. For multiple images illumination normalization, we propose two methods:segmented gamma correction and improved singular value decomposition. Both of them can make the brightness of images close, and the dynamic range of images will coincide mostly. We also presented a quantitative assessment metric of multiple images illumination. It abstracts illumination from images firstly, and then computes the similarity of histograms of the images. Finally we get a parameter between0and1, where1represents the best homogeneity.Experimental results demonstrate that our method can improve the illumination homogeneity among multiple images by subjective assessment and quantitative assessment. We also show statistic data of our experiments, which can prove the effectiveness of our method better.