Multi-scale Region Based Color Correction Algorithm For Stereo Image Pairs

Color correction between two image sequences has become a hot spot of pattern recognition research. It can be applied widely in many fields including the computer vision and remote sensing image fusion. At present, people usually solve the problem of color inconsistency between image sequences of the same environment through global algorithm. But after the correction the result maybe wrong in several local regions. So how to improve the quality of color correction and make the color of images be consistency has been a difficult problem for many researchers.The key points of scale invariant feature transform (SIFT) has the property of remaining unchanged under such conditions as rotation, scaling and affine transform. Also it can process the feature matching between two images which own large color differences without considering such factors as moving object, shading and noise. Based on these advantages, the corresponding relationship of colors between image regions are established and exact information of correction is acquired in this paper.In this paper a new local color correction algorithm based on multi-scale region is presented based on multi-scale concept and methods. At first, the uncorrected image is divided, and then we use the method of feature points matching to acquire suitable correction information. Based on the processing referred above, this paper could implement exact correction.Here are the details of several parts mentioned above:1)In the part of image segmentation, we compare all existing algorithms and choose Mean-Shift algorithm as our method since it could be controlled by simple parameters and the speed is very fast.2)In the part of region corresponding, SIFT feature points matching is adopted to acquire ideal correction information after comparing other color correction methods.3)In the part of correction between regions, due to the poor result that many areas are not corrected after the first correction, we implement the method of region merging based on the information of color texture under the direction of multi-scale concept. And then more correction information is acquired to implement regional correction. This process is repeated until all areas are corrected.The overall experiment shows that the presented method can remarkably improve the accuracy of image color correction comparing with other present methods.