Meng new 1--retinex algorithm for getting started with computer vision

The basic principle of the algorithm is that the human eye view of the image S is composed of two parts, that is, the incident light image L and the reflection image R is multiplied, because R is related to the image contour information, so the problem becomes from S to extract R. The key is to properly model l according to the characteristics of L.

The most basic algorithm thinks that the light is uniform, so the Gaussian filter on the S image, so that the object reflects the information R, and that the object's color is determined by the reflection ability of the object, processing illumination information will not be lost to the color information.

Under the hypothesis of these theories, it is felt that this algorithm is most suitable to deal with the reduction of the object under light, but it is best not to have extra light sources, such as street lamps and building lights, because this will destroy the low-frequency characteristics of L.

Algorithm adopted multi-channel MSRCR algorithm, three-channel standard deviation for the classic 30,150,300, the color has a certain repair.

Example:

Better processing of several images.

　　

The common feature of these is that the observed objects are in the distance, the fog to the object of the degree of ambiguity is more uniform, if the object is near the far, fog only the distant objects blurred, the effect will become worse. Or a light break the original illumination uniformity effect will also become worse (illuminance exists high frequency information).

Poor effect:

　　

　　

We can find that the architectural details in the lower left light area are lost because the algorithm color recovery uses the logarithmic domain linear recovery 128* color specific gravity +127, by the characteristic of the logarithmic function, the area of high brightness is compressed. And then like a street lamp, the brightness of the larger part of the change is smoothed, this is not the brightness of the estimation of the method is improper (this situation should not be considered as a slow change in illuminance).

The algorithm can also solve the situation of insufficient illumination, such as night, early morning, evening pictures, solve the problem that the human eye is not sensitive to the image due to insufficient illumination.

　　

McCann Algorithm

This algorithm is intended to produce a better estimate of the uneven illumination, the extraction of the control information is no longer the weighted mode of the Gaussian convolution, but the intensity of a spiral is selected to be weighted, compared with the Gaussian weighting so that the greater range of illumination information can be obtained, and the specified number of iterations based on the image size.

The details of the algorithm are no longer mentioned, and we look at the difference between MSRCR processing performance.

(MSRCR)

(McCann)

The obvious performance advantage can be seen, but the corresponding computational amount also increases significantly (because of the iterative process).

　　

Meng new 1--retinex algorithm for getting started with computer vision