Tutorial article from ICCV 2019's: "Understanding color & the in- camera image processing pipeline for computer vision", details see here.
Not really translation, interpretation of a lot of hope that we can combine the original pdf screened to see, welcome to ask questions.
It has been introduced the knowledge of the relevant color space , which will introduce a knowledge of other aspects of colors, including white balance and adapt [color] and [with other sRGB color space].
[Color] adaptation and white balance
Said before, we can have three values (e.g., CIE XYZ) used to represent the color, but in the real scene, we see that the color is determined by the reflection characteristics of the object and the scene illumination joint decision, whereas in the previous example we are in default or simply monochromatic white light, which will bring about what impact? Let's look at the following example.
Apple has a figure under different lighting conditions different SPD , also exhibit different colors, but for the viewer in the scene, it was able to see the same red .
As another example, in this picture, the left and right sides coupled with a variety of filters to simulate different light, but although the left and right parts of the roof color makes a difference, we should be able to feel the roof is white (can it?)
The ability to adapt the visual system is referred to as the scene lighting color or chromatic adaptation constant (Color constancy / Adaptation chromatic) . This ability is not perfect, but pretty good with it.
However, the image sensor does not have this ability, in order to remove the influence of the scene illumination, must go through a process step, which is the white balance .
But before that, we need to find a way to describe the light source, thus leads to the color temperature .
The concept comes from the color temperature blackbody radiation theory. This can be simply understood as theory, ideal for a black body, which emitted electromagnetic radiation only, and its temperature dependence .
So we get a map of the temperature (in Kelvin units) are mapped to the electromagnetic spectrum of black body radiation , while the electromagnetic spectrum in the region of visible light appears to effect field can be used to describe a light source .
Here refers to the temperature of blackbody temperature, instead of the light source, so compared to the temperature, in fact, we should say ** CCT (Correlated Color Temperature, CCT) ** expressed to distinguish.
The above image shows the table and the color temperature of the light source, there may be some counter-intuitive that the lower the temperature but look warmer , then again, where the temperature refers to the temperature of an ideal black body, not your heart temperature \ manual doge.
For convenience of description, the CIE established on behalf of several synthetic SPD as the real source, number listed below:
A Tungsten
B midday sun
C average day sunlight
Natural daylight (5000K, 5500K, 6500K) representative D at different color temperatures, generally written D50, D55, D65
E ideal light source having a constant energy such SPD and does not represent any real sources, but the D55 and the like
F Series: Analog various fluorescent lamps (total of 12)
when we buy bulbs may see these parameters.
In the color space, that the light source SPD corresponds to a white point , the nature of the chromatic adaptation is to different scene illumination white point become the same.
This picture, the colors on the curve are in fact a scenario where "white."
Next, look at a simple color constancy process (Von Kries Transform) . (LMS) 1 and (LMS) 2 refer to different scenarios three kinds of cones stimulus values , and the subscripts W under this scenario means that white point corresponding to the stimulation values .
Thus, we put the converted tristimulus values of the scene 1 to the scene 2, thereby to eliminate the influence of light. Look at an example.
We can see, we first define a scene in white, and then they adjusted to be the same, so the color looks different scenarios more harmonious.
[With other sRGB color space]
Okay, so far we do not get the color?
Fast fast (really).
Although the CIE XYZ color space is an authoritative, but the image and equipment rarely used directly XYZ. This is mainly because XYZ color itself does not represent , we left a lot of space in the industrial or prefer to use the RGB color space, although they can not represent all the colors, but it is easy to control and understand. In order to allow equipment from different vendors to have a unified standard, so the sRGB born.
In 1996, Hewlett-Packard, Microsoft, and defines a series of RGB primary colors:
R & lt the xyY the CIE = (0.64, 0.33, 0.2126)
G = the xyY the CIE (0.30, 0.60, .7153)
B = the xyY the CIE (0.15, 0.06, 0.0721)
They thought it was the most devices can achieve RGB color space.
The white point is set to a D65 illuminant.
Note that the specified white point is an important thing , which means that sRGB is the assumption that the viewing conditions established under the premise of (6500K daylight). Whenever we have a map to the CIE XYZ color space, you need to specify the white point .
The figure is mapped sRGB effect on the CIE xy chromaticity diagram (but said before, is not recommended xy point of view in CIE ...).
The figure is the CIE XYZ linear sRGB conversion matrix (negative discard the process).
Why here say linear? Because to this step is not enough, from the linear sRGB to sRGB another step gamma correction .
What is gamma corrected it?
You can see an example of this:
There are two maps, each map there are two blocks of color on what you feel Zhang figure is half the color brightness of color?
This example is from https://www.cambridgeincolour.com/tutorials/gamma-correction.htm, there is a detailed description of the gamma correction on the page, simply, the reason why there is a gamma transformation, because the human eye to light changes in brightness perception is non-linear in. And the camera (photoreceptors) compared to our light than in the dark more sensitive to changes in light, which allows us to get some protection in the event of outdoor light. The purpose of the display is to reproduce human experience, so to be adjusted.
A figure is just after the gamma correction, so that A exhibits FIG eyes half luminance, and B is shown in FIG physical half intensity. (And do you feel the same?)
Formula called Stevens' Power LAW , is the model used to describe this phenomenon. I is the input image of the stimulus value, S is the after treatment, consistent with human feelings stimulus value image. When taking a positive number less than 1, that is, the color space of the camera to the human feelings conversion (encoding) when a fetch is greater than 1, which is the inverse transformation (decoding). Different models with different coefficients have values in the sRGB, a is taken 2.2.
In addition to sRGB, as well as NTSC, Adobe RGB color space and other kinds of standards, their scope vary, but they all come from CIE XYZ space.
Note that there's a "Y" are not the same.
It later appeared in a variety of other color spaces, they are designed for different purposes.
CIE LAB color space
This color space is perceived in order to get a uniform color space. While the CIE XYZ space provides a mapping between SPD and radiation color perception, in CIE XYZ space, but the change is not uniform perceptually uniform variations of. As shown, each ellipse figure shows human eye experiences a similar region (enlarged ellipse).
Secret
fear
police
report
The CIE CIE Lab color and converted into a change in luminance more uniform color space, L represents lightness, a and b represent hue plane spanned . The following figure shows the colors on the ab plane.
Y'UV Y'IQ Y'CrCb color space
These spaces are the RGB space is decomposed into " Class luminance " composition and color composition, noted that these color spaces of Y is not defined in the linear-sRGB or in linear-NTSC , they define the gamma encoding the sRGB and NTSC color space. Normally it should be written Y 'in order to avoid misunderstandings, but generally will be written Y. (The YUV, YIQ, YCrCb Y in CIE XYZ as the Conference on Computer Vision Y is a common mistake). Such space can be discarded directly to obtain the U and V image gray levels, may be used for monochrome and color television signals in common.
Finally, to introduce color .
Based on the CIE 2000 CIE LAB color error matrix is defined, it returns an error in color 0-100. It will also be referred to as CIE DE2000, CIE DE, ΔE, Delta E, DE. Reference scale errors as follows:
Generally 2 or less color is better, which means that unless a standard observer observe very closely, otherwise it is impossible to distinguish between two colors.
it is good!
Color crash course is over, then we can formally study the camera!
