Edge detection and connection of image segmentation realized by MATLAB

Edge detection and connection of image segmentation

One 、 Implementation details

This section is divided into four parts to explain the implementation details , Corresponding to this operation ： Three methods of edge detection and edge connection

1.1 Sobel Basic edge detection

Implementation steps

Problems encountered

The selection of threshold needs to be adjusted artificially , Too low threshold makes it difficult to remove noise points , Too high threshold makes the edge discontinuous , The final selection threshold is 30%; Besides , While thresholding , Binarization is necessary , Facilitate the subsequent processing of edge connection .

1.2 Marr-Hildreth Edge detector

Implementation steps

Smooth the image with Gauss low-pass filter

Calculate the Laplace of the result of the previous step

Find the zero crossing of the above results

Problems encountered

On the calculation of zero crossing , It was originally written according to the courseware , Ask up and down 、 about 、 The two diagonally opposite corners have a signed difference greater than the threshold value of zero crossing , But the effect is not ideal , There are many burrs on the edge . Later, I referred to an online blog , See the end of the report for specific links , Improved zero crossing Algorithm ： First the 3*3 Area passing 2×2 Of all the 1 Filter convolution , Generate 2×2 Region , Then judge whether the maximum and minimum values in the region have opposite symbols and are greater than the threshold .

1.3 Canny Edge detector

Non maximum inhibition , That is to find the direction closest to the gradient direction , If the current pixel is not the maximum value in the direction, the hysteresis threshold is suppressed , That is, two thresholds are used to divide the pixel set into 2 Partial connectivity analysis , The low threshold set will be denoised , Remove all points that are not connected to the high threshold set , Then merge the two sets

Problems encountered

The writing method of non maximum suppression according to the courseware is to set the non maximum pixels to zero , Considering the efficiency of the actual code , Reverse operation , Because the image is sparse , So first set the whole to zero , If the pixel is the largest , The pixel value is set to 255

1.4 Edge connection method

Implementation steps

Given the binary image and starting point, if the starting point pixel is 0 And end it , Otherwise, add the starting point output For the starting point 8 Neighborhood , If there is a point nearby that has not been visited , And the pixel value of the point is 255, Then add to output The newly added point is the new starting point , Cycle according to paragraph 3 Step by step to find .

When 8 When there is no qualified point in the neighborhood , Stop the cycle . Output output

Problems encountered

It was originally intended to use gradients when judging local similarity , But the effect is not ideal , So simply deal with , That is, find it directly 8 The pixel value in the neighborhood is not 0 Just a few points , Of course, you need to avoid repetition when looking , The essential idea is graph traversal ;

Special treatment is required for boundary conditions , For example, the points on the edges and corners of the whole image cannot be used directly 8 Connectivity , Otherwise, the array is out of bounds ;

Pay attention to the order of coordinate parameter transmission , imtool The parameter order of the function is listed first (x) Again (y). And what we achieved my_edgelinking The order of parameter passing is first (row) Then list (col).

The starting point is to join output after , Do not mark it as arrived , The purpose is to facilitate reaching again in the future and make the boundary become a closed curve

The method of edge detection has the best effect Canny detector , On this basis, continue boundary tracking

Two 、 result

This section is divided into three parts , Introduce the structure of the experiment 、 Experimental results 、 Conclusion analysis

2.1 Experimental setup

code The corresponding frame code in the file edge_test.m、my_edge.m、my_edgelinking.m.

among my_edge The parameters of the function method by 1、2、3 Time corresponds to sobel、Marr-Hildreth、Canny Three methods

asset\image The corresponding picture file in the folder

Run the code directly to view the original image 、 The effect of various edge detection and edge connection , To replace the input image, put the imread Function parameters can be modified .

doc In the folder is the report you are reading

2.2 experimental result

2.2.1 rubberband_cap.png

2.2.2 ayu.jpg

2.2.3 giraffe.jpg

2.2.4 leaf.jpg

2.2.5 noise.jpg