Image features and extraction

This paper is mainly used to record image features and their extraction methods

---

1. Color features

     Color feature is a global feature , Describes the surface properties of the image or the corresponding scene in the image area . General color features are based on pixel features , All pixels contribute to the image or image area . Due to the direction of color to the image 、 Size and other changes are not sensitive , Cannot capture the local characteristics of the object well .
    Express the color characteristics , Its essence is to divide the color into different intervals , That is to quantify the color . Methods of characterizing color characteristics , Here are two kinds of , Then there are quantization color histogram and clustering color histogram .

1.1 Quantify color histogram

Applicable color space ：RGB、HS Color space
1.RGB： Based on the three basic colors of red, green and blue, they are superimposed to varying degrees , Produce a rich and wide range of colors , Therefore, it is commonly known as the three primary color model .

2.HSV：HSV(Hue, Saturation, Value) According to the intuitive characteristics of color A. R. Smith stay 1978 A color space created in , Also known as the hexagonal cone model (Hexcone Model). The parameters of color in this model are hue (H)、 saturation (S) And lightness (V).
The specific method of quantifying color histogram ： Divide the image into several small units , The color of the center of each small cell represents the color of this cell , Count the number of pixels falling on the quantization unit .
advantage ： Computational efficiency
Inferiority ： Quantization problem and histogram sparsity

1.2 Clustering color histogram

Applicable color space ：Lab Color space
lab：Lab The color model is made up of brightness (L) And about color a, b Three elements make up .L Express lightness (Luminosity),a Indicates the range from magenta to green ,b It means the range from yellow to blue .L The value range of is determined by 0 To 100,L=50 when , Equivalent to 50% Black ;a and b The range of values is determined by +127 to -128, among +127 a It's red , Gradually, the transition to -128 a When it's time to turn green ; Same principle ,+127 b It's yellow ,-128 b It's blue . All colors are composed of these three values . for example , A piece of color Lab The value is L = 100,a = 30, b = 0, This color is pink .

Clustering color histograms ： Use the clustering algorithm to cluster the color vectors of all pixels , The color of each small cell is represented by the color of the cluster center . This histogram avoids the problem of sparse histogram , That is, the color displayed in the histogram is the color used , Other colors don't take up space , Will not show .

2. Geometric features

2.1 Edge features

Edge features ： Areas where pixels change significantly , It has rich semantic information , It can be used for object recognition and geometry 、 Change of perspective .
A mathematical description of edges ： The extreme region of the first derivative .

For vertical edges , The first derivative can be directly used to distinguish .
But in fact, the first derivative of the image , It's not as smooth as the picture （ There is noise in the image ）. In order to denoise , Gaussian smoothing can be used .
Suppose the original image function is f, The Gaussian function is g, Then the new function obtained by Gaussian smoothing h=f*g.
For new functions h To find the derivative , You can get smoother edge features .

For sloping edges , You need to find the gradient of the function . The position where the edge exists is the direction in which the gradient increases fastest

2.2 Feature descriptors based on feature points

For the same object , From different distances 、 Direction 、 angle 、 When observing under illumination , The size of the object 、 shape 、 Light and dark are different , But we can still judge that it is the same object . In order for computers to have this ability , A feature descriptor based on feature points is introduced . The ideal feature descriptor should have the following properties ： In size 、 Direction 、 In different light and shade images , The same feature point should have enough similar descriptors , It is called the reproducibility of descriptors .
In order to realize this feature descriptor , We need to find the key points in the picture . For a picture of a dog , Constantly look from the front or the side , It has some characteristics that are different from people , Depending on the feature points, we can distinguish whether the picture is a dog or a person . The most important feature of stabilizing local feature points is to resist image changes .
For the same place , Take pictures from two different angles , By finding feature points, we can realize the splicing of pictures

2.2.1 Look for feature points ：Harris Corner detection

Corners are a kind of remarkable feature points . Use a small observation window , In the process of moving the observation window in the picture , As long as you encounter corners , There must be large pixel changes .

Its mathematical model is described as ：

Approximate the formula , You can get ：

Among them M：

The two terms in the matrix are called eigenvalues
If the image is a straight line , Then an eigenvalue is large , An eigenvalue is small
Screen in image , Then both eigenvalues are small , And it's almost the same
Corners in the image ： Both characteristic values are large , And it's almost the same

2.2.2 Another feature point ： speckle

For the edge mathematical model just mentioned , The place where the extreme point of the first derivative exists is the edge . Where there is a maximum point for the second derivative , That is, spots .

2.2.3SFIT： Characteristic descriptors of spots

computational procedure ：

• Obtain extreme points in the differential Gaussian scale space
• Handle key points , Including position interpolation and edge point removal
• Estimate the direction of key points
• Rotate the regional coordinates
• Calculate histogram of sampling area
• Generate key point descriptors

Gauss difference ： Two Gauss check images with different scales are used for processing , Then subtract the processed results . After subtraction, you can get the edge in the image . The specific method is Gaussian pyramid ：


Make a difference between the two layers in turn , Then we can get the Gaussian difference space

SIFT Characteristics ： It has good invariance ; Unique , Rich in information ; It has many quantities