Computer Vision No. 1. Image Preprocessing



1. Color space

  • RGB is the more superimposed, the brighter (usually used for computer display)
  • CMY is the more superimposed, the blacker (usually used for printing)

Insert picture description here
Insert picture description here


Conversion formula of gray scale image:
Insert picture description here


2. Histogram equalization

Non-linearly stretch the image. Redistributing the number of pixels in each gray scale unit is generally for the whole picture taken as too white or too dark. If it is partial darkness or volume, then it is fine. An adaptive histogram can be used.


3. Morphology

Insert picture description here


Insert picture description here


4. Filtering/Convolution

Different functions need to define different functions

  • Smoothing, denoising
  • Gradient, sharpen
  • Edge, salient point, texture
  • Pattern detection

4.1 Filling before filtering

To keep the size unchanged, the filling operation before filtering:

  • Zero padding
  • Boundary copy
  • Mirror image
  • Block copy

The specific effect is as follows:
Insert picture description here
Insert picture description here
At present, the zero-padding method should be used the most.



4.2 Several classic filters

  • Average filtering: generally not used much, that is, it cannot denoise, nor can it extract features well.

  • Median filtering: generally used for denoising.

  • Gaussian Wave Recorder: Blur the surroundings, a bit similar to the background blur function of a mobile phone camera. (The main function is to build a Gaussian pyramid)

  • Prewitt filter for gradient detection:


Insert picture description here

Insert picture description here


4.2.1 Laplacian 算子

The above methods to find the gradient are equivalent to the gradient of the first derivative, and the laplacian operator is the second derivative. Note that all his convolutions and sums are equal to zero.


Insert picture description here

The laplacian operator, because the total sum is equal to 0, when the picture is very uniform, it is 0. When a certain pixel is more prominent, the final filtering result will be more prominent, then a sharpening can be achieved The process, below is his renderings.

Insert picture description here
In addition, there is a simple filter to achieve the sharpening effect.

Insert picture description here




5. Image pyramid

There are two general image pyramids, one is the Gaussian pyramid, and the other is the Laplacian pyramid.

5.1 Gaussian Pyramid

Ordinary pictures, if they are directly down-sampled, a lot of details will be lost. You can pass Gaussian filtering first, and then down-sample the pictures after Gaussian filtering, and the resulting pictures will be clearer. As shown in the figure below: The image pyramid formed by the Gaussian filtering method is the Gaussian pyramid.

Insert picture description here



5.2 Laplace Pyramid

The purpose of the Laplacian Pyramid is to make up the missing part after the down-sampling of the Gaussian pyramid is completed, and then the up-sampling restoration. To put it bluntly, he is a difference.

Insert picture description here


Insert picture description here



Digression

1. Why are filters generally symmetrical?

In fact, the main purpose is to reduce the complexity of the calculation.

Insert picture description here



2. Curtain

Jiang Wei reduces the dimensionality, and then removes some less important data.

Guess you like

Origin blog.csdn.net/zhaozhao236/article/details/109485311