I refer to the paper here : Local Consistency Index Scheme of Random Valued Impulse Noise Detection System
Article Directory
Preface
This blog mainly describes how to use the LCI value to determine whether the pixel is a clean pixel. The first time the writing is wrong, please advise us.
1. What is noise detection?
Due to non-ideal sensors, transmission errors, wrong memory, etc., the luminosity value of some pixels in the digital image may not be correctly evaluated. These damaged pixels will affect the overall image vision, so we need to detect the forced noise pixels Use this to lay the foundation for the restorative images.
Second, the definition of LCI value
1. The characteristics of noise pixels and clean pixels
The special part c is the rough area
According to the picture, we can easily get the following conclusions:
1. Regardless of whether it is located in a flat area or a rough area, for a pulse pixel (noise pixels are divided into two types of fixed value pulses and random value pulses, we are mainly talking about random pulses here) Usually you can find pixels with similar intensities and clean pixels usually have a certain amount of similar intensities
2. We can use the number of similar pixel values near a given pixel to determine whether the pixel is an impulse value
2. How to determine the value of LCI
This LCI value can be understood as the similarity of a pixel to the entire picture. Of course, the larger the LCI value, the higher the probability that the pixel is a clean pixel. So how to define this value?
Where λ (x, y) \lambda(x,y)λ ( x ,y ) uses Euclidean distanceS (x, y) \mathit(S)(x,y)for geometric distance
S(x,y ) is the pixel intensity.
It is obvious from the above that the similarity is not only determined by the geometric distance or the intensity. After that, we will add the appropriate weight to the geometric distance to calculate the LCI value.
In fact, we can compare at this step, but because the calculated value has a large deviation in each value, we averaged it so that its value was around [0, 2.5] and then normalized it to limit the value to [0, 1]
However, after later experiments, it is found that the detection accuracy of values that are not normalized is actually higher (the specific reason is that I don’t think it is rigorous enough after understanding, so I won’t repeat the
experimental results here . We set the threshold to 0.5 but in fact There are still cases of false hits and missed detections, so we will continue to improve
the formula analysis in the follow-up . The threshold is actually a criterion for whether the detector can accurately detect noise. If the selection of the threshold is biased, it will cause a series of problems such as a decrease in accuracy. And through the above pictures, we can actually know that there is a big difference between the LCI value of the flat area and the rough area. In order not to cause a decrease in accuracy, we need to set two thresholds, one for the flat area and the other for the rough area
. I will describe in detail in the next article
to sum up
1. The LCI value can be used to judge whether a pixel is a noise point
. 2. Judging whether a pixel is a noise point depends not only on the intensity of the pixel, but also on the geometric distance. There is no much reference for judging the similarity of pixels. value