Retinex theory is the color of the last century seventies proposed by the Land. I think the core idea based on two points
(1) when the color perception, the human eye local absolute intensity relative intensity superior sensitivity.
Real pattern (2) reflection component R (x, y) stored in the presence or absence of the object light, a light image of an estimate of their component L (x, y), the more accurate the obtained R (x, y) the more accurate.
First will be described a first point, human perception of color is not based on absolute light intensity, the reaction can be understood on the image, the color perception of the human eye to a pixel (X0, Y0) which is not based on RGB three the absolute value of the channel, but associated with two factors.
1, three-channel points of difference values (X0, Y0), it will cause a color shift.
2, (X0, Y0) of the remaining pixels of a neighborhood of RGB values, which will form a perception of color, the absolute value of the leading three-channel RGB brightness sensation.
I can not present quantitative and qualitative analysis of their relationship, because it involves the feeling that mysterious thing, but we will use on the line.
Again that a second point, when a given scene, if no light within the scene, then R (x, y) should be a constant. Whether light L (x, y) to change how, R (x, y) of the same. In other words, no matter what color what my light intensity, which I should be able to be corrected to standard lighting conditions. This is also reflected in the color constancy.
So the question is, based on the above reasoning how to do to correct the multi-light image.
Retinex algorithm for development, from the initial random path method, to McCann iterative method, variational method and then to surround the center. Their explanation about the Internet to find a whole lot, I will not repeat them here.
Let's start with the single-scale-Retinex, SSR algorithm can be found online, said the effect of poor basic truth, not simply reflect the characteristics of its color constancy algorithms. In fact, if you had access to large amounts of data, you'll find the original text of the SSR NASA explained unclear, and online codes can say it is a non-recurring reference, you may even find that they surround function described They are inconsistent. SSR final analysis is that local illumination change, a weight value in the neighborhood of the illumination is estimated.
SSR will usually choose Gaussian function or Gaussian function, you can use matlab in fspesical function, you can also write your own directly, in fact, not a big problem. General first met SSR most puzzling is how to scale what is clear, when you look to achieve several SSR, you will find more and more doubts. How to determine the scale, in the end, and the number and dimensions of the relationship between the sigma value of the Gaussian function, and finally in the logarithmic domain need to return to the index domain, or need normalization, the final image color cast how to solve the problem. The details of the deal will significantly affect the picture quality, especially visual effects.
Developed on the basis of the MSR and SSR MSRCR, but to be honest, because their implementation is based on SSR, the SSR so if you have not written, so they do not achieve a certain satisfactory.
On the basis of the MSR, in order to achieve true color constancy calculation, I made changes to MSR.
Step roughly divided into three steps:
1, instead of using a Gaussian filter filtering guide, the conventional MSR different algorithms, but this three-channel regardless of channel processing unity estimated illumination. This is based on a cognitive, light is continuous in nature, the RGB three-channel display screen for convenience only, three separate channels formed by processing.
2, the light of the estimated adaptively adjusted, the reflection amount is calculated in the logarithmic domain r (x, y)
3, r (x, y) to be adjusted in order to adapt to different environments, I used a little trick to achieve a color fidelity, fidelity optional whether in the algorithm, the fidelity of the original premise is good enough quality, image elements are abundant, if not maybe a prerequisite when it will calculate the true color of a serious error. As to why so, which is the essence of Retinex algorithm is related to light because of the more accurate the estimate, the correction to the quality standard light source, the better. The Retinex illumination is based on an estimate of the spatial neighborhood.
Note: 2 is achieved in the logarithmic domain, which is similar to the MSR
FIG attached explanation of the effect of some
The origin of the artwork, Southeast University, a friend offer
Fidelity is not calculated, it can be seen that the color constancy reflect poor
Conduct fidelity, fully reflect the color constancy
Original, also provided by the Southeast University friends
Fidelity does not compute
Fidelity calculated
Original to the network
No fidelity, visible light is red
After treatment, magenta light is weakened, but FIG visible light is amplified, and because the picture quality is not high, so that the image quality is calculated fidelity visual fidelity is not lower than
These are my superficial knowledge of the Retinex.
November 7, 2019 at Beijing University of Aeronautics and Astronautics
Automation Science and Electrical Engineering, Fu Yun Song
Contact: [email protected]