Algorithm principle
These three methods should be relatively basic color equalization processing algorithms.
The principles of the three algorithms are relatively simple. You can Baidu for the specific principles.
(1) The wallis color equalization principle mainly relies on using the Wallis filter to make the mean and standard deviation of the original image consistent with the reference Images are matched so that the original image and the reference image have similar tones.
(2) Histogram matching is to use the cumulative distribution function to change the histogram of the original image so that it is similar to the histogram shape of the reference image, thereby achieving the purpose of color uniformity.
(3) Color transfer was initially performed in the Lαβ color space by adjusting the mean standard deviation. Later, some people proposed color transfer in related color spaces, especially in the commonly used RGB color space. The essence of this method is to use scaling, rotation and Translate to move the data points of the original image to adapt to the data point cluster of the reference image in the RGB color space (this data point cluster is the three-dimensional data point cluster in the previous article). Personally, I think this is an evolution from two-dimensional (histogram matching is to match the shape of the histogram) to three-dimensional (matching of data point clusters in RGB three-dimensional color space). The color transfer mentioned below is all the shape of data point clusters in RGB color space. Location match.
Comparison of main processing results
(1) Let’s look at the more normal ones first.
For this kind of ground objects that are less complex and more evenly distributed, these three methods can almost make the original image and the reference image have similar tones.
(2) This image with highlighted buildings looks like the color transfer technology is closer to the ref reference image, but the buildings inside are slightly brighter.
(3) The background in the ref reference image changes to blue. The histogram matching result is distorted. Wallis looks better, but there is still a gap with ref. The color transfer is also different from ref, but it is much better in comparison.
Summarize
(1) Generally speaking, wallis adjusts the mean standard deviation of an area. When there are features with large mean contrast in the area, the wallis processing effect will be correspondingly worse (so the outlook direction in many uniform color papers is It fails to take into account special areas such as the shadows of buildings in highlighted waters. These areas should be identified first and then processed separately. This is reflected in Wu Dajiao's software GeoDoging) Global wallis is suitable for processing images with relatively uniform distribution of ground objects (2)
Histogram Matching generally involves stretching the histogram shape. Although some make the grayscale range larger, it is easy to lose the intermediate grayscales and cause distortion.
(3) The color transfer method based on RGB related color space is generally effective, but there is no one. The method is universal (one reference image is a Wuhan city building image with red buildings + green space. The original image has a layer of shadow-like shadows on it. The color transfer processing effect of the lower brightness image is not good. Maybe the difference between the two images is too big. The matching effect of the three-dimensional scatter plot is not good)
