CV learning notes - edge extraction

Edge extraction

1. summary

1. edge

Definition ： The edge of the image refers to the part where the brightness of the local area of the image changes significantly , The gray profile of this area can generally be regarded as a step , That is, from a gray value that changes sharply in a small buffer area to another gray value with a large difference in gray level .

characteristic ： There are positive and negative edges , From dark to light is positive , From light to dark is negative .

Algorithm for calculating edge amplitude ：sobel、Roberts、prewitt、Laplacian、Canny operator (Canny The effect of operator is better than others , But it is more complicated )

2. Definition of edge extraction

Edge detection is mainly the measurement of the gray change of the image 、 Detection and location

3. Application of edge extraction

For example, semantic segmentation 、 Instance segmentation

2. principle

1. high frequency & Low frequency signal

The low-frequency and high-frequency signals in the image are also called low-frequency components and high-frequency components . High frequency component refers to image intensity in the sense of image （ brightness / Grayscale ） Where there are drastic changes , That is, the edge （ outline ）; The low-frequency component in the image refers to the image intensity in the sense of image （ brightness / Grayscale ） Where the change is gentle , It is characterized by large areas of color patches .

notes ： Human eyes are more sensitive to areas of high-frequency signals

2. Principle and steps of edge detection

principle ： At the edge , Pixel values appear ” jumping “ Or a big change . If you find the first derivative at this edge , You'll see extreme values appear , And where the first derivative is the extreme value , The second derivative is 0, Based on this principle , Edge detection can be carried out . And in the image , It's a two-dimensional space , Two dimensional space requires derivation in two directions , The extreme value can be found by locating the pixel whose gradient value is greater than the neighborhood （ Or it can be extended to more than a threshold ）.

3. Common filter operators

• Sharpen the image ：（ Laplace transform kernel function ）, Image sharpening is to compensate the contour of the image , Enhance the edge of the image and the gray jump part , Make the image clear . Image sharpening is to highlight the edge of the object on the image 、 outline 、 Or the characteristics of some linear objective elements . This filtering method improves the contrast between the object edge and the surrounding pixels , So it's also called edge enhancement .

  

• Image smoothing ： Image smoothing refers to a wide area used to highlight an image 、 Low frequency components 、 Trunk components or image processing methods that suppress image noise and interfere with high-frequency components , The purpose is to make the brightness of the image smooth and gradual , Reduce the mutation gradient , Improve image quality .

• Sobel operator ：

  Definition ： Contains two groups 3*3 Matrix , Transverse and longitudinal formwork respectively , Convolute it with the image , Then we can get the approximate values of the horizontal and longitudinal brightness difference respectively . In practice , The following two templates are commonly used to detect image edges ：

​ advantage ：Sobel Operator is a typical first derivative edge detection operator , Due to the introduction of an operation similar to local average , Therefore, it can smooth the noise , It can eliminate the influence of noise , The influence of pixel position is weighted , Therefore, Prewitt Better than operator .

​ shortcoming ：Sobel The operator does not process based on the gray level of the image , Therefore, the theme of the image cannot be strictly separated from the background , because Sobel Operators do not strictly simulate human visual physiological characteristics , So sometimes the extracted image contour is not satisfactory .

• Prewitt operator

  Prewitt Operator is a kind of edge detection of first order differential operator , Use pixels up and down 、 The gray difference between the left and right adjacent points , Reach the extreme detection edge at the edge , Remove some false edges , It can smooth the noise . The principle is to use two direction templates to convolute the image neighborhood in the image space , These two direction templates detect a horizontal edge , A detection of vertical edges .

4. Canny Edge detection algorithm

1. Grayscale the image

Process the image into gray image .

2. Gaussian filter the image

According to the gray values of the pixels to be filtered and their neighborhood points, the weighted average is carried out according to certain parameter rules . This can effectively filter the high-frequency noise superimposed in the ideal image .

Gaussian smoothing presents Gaussian distribution in horizontal and vertical directions （ A two-dimensional ）, It has better smoothing effect than mean filtering .

A one-dimensional 、 Two dimensional Gaussian probability density function ：

The choice of the size of Gaussian convolution kernel will affect Canny The performance of the detector ： The bigger the size , The lower the sensitivity of the detector to noise , However, the positioning error of edge detection will also increase slightly .

notes ： Generally, the actual choice 5*5, In most scenes, the comprehensive effect is the best .

example

3. Detect the level in the image 、 Vertical and diagonal edges （ Such as Prewitt、Sobel Operators, etc ）

See the previous section “ Common filter operators ”

4. Non maximum suppression of gradient amplitude value

Non maximum suppression （NMS）： Search for local maximum , Suppress non maxima .

Purpose ： Eliminate redundant boxes

Method ： For overlapping candidate boxes , Calculate their overlap , If greater than the specified threshold , Delete ; Below the threshold, keep . For candidate boxes without overlap , All reserved .

• Compare the gradient intensity of the current pixel with two pixels along the positive and negative gradient direction .
• If the gradient intensity of the current pixel is the largest compared with the other two pixels , Then the pixel point remains as the edge point , Otherwise, the pixel will be suppressed （ Set to gray value 0）.

Example ： Take target detection as an example ： In the process of target detection, it is likely to produce some overlapping candidate boxes at the position of the same target , At this point, you need to find the most appropriate target bounding box , Non maximum suppression is a good solution to eliminate redundant bounding boxes .

5. Using double threshold algorithm to detect and connect edges

background ： The output result of the image after non maximum suppression processing is a non edge point with a gray value of 0, The local gray maximum point that may be the edge is non-zero （ Usually take 128） The binary image of . Many false edges caused by noise or other reasons are retained in the result . Based on the background of this problem , Double threshold detection is proposed .

Double threshold detection ：

• If the gradient value of the edge pixel is higher than the threshold , Then mark it as a strong edge pixel ;
• If the gradient value of the edge pixel is lower than the high threshold and higher than the low threshold , Then mark it as a weak edge pixel ;
• If the gradient value of the edge pixel is lower than the low threshold , It will be inhibited .

The idea of double threshold detection ： Above the high threshold is a strong edge , Below the low threshold is not an edge , Between the middle is the weak edge .

notes ： The choice of threshold depends on the content of a particular image , There is no universally applicable threshold .

Suppress isolated threshold points ：

For weak edges , It can be either real edges or retained by noise or color changes , therefore , The weak edges introduced by the latter case should be suppressed , Suppress inappropriate weak edges based on the following two theoretical steps .

• In general , Weak edges introduced by real edges will be connected to strong edge pixels , The noise response is not connected
• To track edge connections , By looking at the weak edge pixels and their 8 Neighborhood pixels , As long as one of them is a strong edge pixel , Then the weak edge point can be kept as the real edge .

---

Personal study notes , Learning exchange only , Reprint please indicate the source ！