Study On Color Harmonization Of Remote Sensing Images

Nowadays, large-scale high-resolution remote sensing images have been widely used in many fields, such as resource statistics, geographic information system (GIS), geologic research, military affair and national defense, disaster estimation, environment supervision and etc.. But because of the differences between raw sub-region image data taken by variant satellites under changefully weather at different time, great gaps may arise in the color of the whole image after image mosaic, which leads to a lot of difficulties in later work such as analyzing geographic data and decreases the accuracy of the analysis. In practice, software such as Photoshop is usually used to harmonize the whole color of the image, which are often tedious and of great human labors as well as time. In this paper, we introduce a new color harmonization algorithm to enhance the harmony among the colors of a given image. Given an un-harmonious image, our algorithm automatically enhance the color "look-and-feel", only with a few user interactions, moreover, the main idea of our algorithm can be extended to other image processing fields.This paper is organized as follows. Chapter 1 introduces the basic ideas of remote sensing images as well as their processing methods, reviews previous related work in this area, and briefly describes our main research work of this paper.In Chapter 2, we describe our algorithm for harmonizing two colors. First, we probabilistic segment the source and image based on K-means partition and Expectation-Maximization (EM) method respectively. Second, we sample from each region of the source image. Third, for each pixel in the target image in scan-line order the best matching sample point is selected based on the weighted sum of luminance and previous probabilistic segment result. Once the best matching pixel is found, the chromaticity values are transferred to the target pixel while the original luminance value is retained.In Chapter 3, we present a method to harmonize several colors. Based on the algorithm in Chapter 2, the user interactively specifies different source images for different regions for the sake of harmonization, then, Graph Cut is performed to get the seamless compositing result.In Chapter 4, we extend the main idea of our harmonization framework to grey-scale image colorization, the natural color generation of Infrared thermal images and multi-spectral fused images and converting visible images into infrared images based on thermal database of materials.The summary and conclusion of this paper are given in Chapter 5, as well as the discussion of the further research work.