Skip to content

Latest commit

 

History

History
272 lines (178 loc) · 15.8 KB

File metadata and controls

272 lines (178 loc) · 15.8 KB

Polyester-Yarn-Microscopic-Image-Processing-and-Data-Analysis

Give a roughly or manually measurement of fibers from the microscopic image of the cross section of yarn (This application requires .NET 3.5/2.0/3.0 installed)

GIF 1. All the function are demonstrated here.

Function of the Project

Auto Detection

  • Detect each fiber in the yarn automatically and give a roughly measurement of the diameter (unit in pixel) in the cross section of the fiber.

Image 1. Circles auto detection

GIF 2. Auto Detection.

There are four parameters which require specify by the user.

Min R and Max R are for definding the searching space.

PixThred is the threshold of pixels value which the user want to detect (The value ranges from 0 to 255. RGB value are the same because we are analyzing the grayscale image).

Accuracy is to define what the criteria is to define a circle. The smaller the accuracy, the application will be more susceptable to the noise and give some circle with only a few white dot on that circle. However, if you set the accuracy too high, the application will hardly give all the circle detection.

Small is the ratio compares with the largest circle (marked in blue) which the application can detect in the image. If any circle compares with the largest one smaller than the ratio, will be marked in red. Other circle will be marked in green.

Manual Measurement

  • Provide an user interface to manually measure the width in the cross section of the fiber.

Image 2. Circles manual detection

About the Theory

The gist of this application is circular object detection. The more percise term is Circle Hough Transform. First, we need to make the edge more clear from the image. Second, we need to detect the circle.

The whole process can be summrized into the following flow chart. Each step will be specified in the following paragraph. Notice that after the edge detection, the image will be devided into 3 parts so that we can process these three parts at the same time. Not untill these three parts are finished will us collect all the detected circles from them and then go through the merge process.

Image 3. Flow chart of the algorithm

Image Edge Detection

There are various kind of edge detection algorithm such as Laplacian, Sobel, Kirsch and Prewitt. You can refer to this article by this link : https://softwarebydefault.com/2013/05/11/image-edge-detection/. We use two files Matrix.cs and ExtBitmap.cs provided by the website which I just mentioned. Some basic convolution method implemented in the edge detection can refer to this video: https://www.youtube.com/watch?v=XuD4C8vJzEQ and the following GIF from OpenCV.

GIF 3. Convolution (By OpenCV)

In our application, we first apply kirsh edge detection on the selected image, and then we apply the Sobel edge detection on the image which has been processed by the kirsh edge detection.

Image 4. Image processing from the original image to gray style with edge detection

Circle Detection For Unknown Radius

The application will examine the pixel every 4 columns for every 4 rows in the image showing in the following GIF. The pixel marked with yellow will be regarded as the center of a circle.

GIF 4. The examine steps of the image (also see GIF 7).

We make a simple GIF in the following to demonstrate the algorithm of the Circle Hough Transform. In the GIF, we first determine the pixel which is marked as yellow dot. We make various radius of circles based on that yellow dot.

We've already have some basic knowledge on the dimension of the circle in the given image, so we can narrow down the radius searching space for example from r = 10 to r = 15 (unit in pixel). The user can enter the MinR and MaxR to set the searching space. We will detect how many pixels which the pixel value (either Green or Red or Blue value) is larger than a given threshold on the virtual circle at given radius (we examine the pixel value every 3 degree on the virtual circle to reduce the calculation time). The number of pixels (vote) should also be larger than a given threshold.

GIF 5. Circle detects based on certain pixel as the center of virtual circle

In the above example, for yellow dot as the center of virtual circle, we can see r = 10 (count=0), r = 11 (count=0), r = 12 (count=0), r = 13 (count=20), r = 14 (count=38), r = 15 (count=100). If we set the count (vote) threshold = 30, then only r = 14 and r = 15 will be taken into decision. We select the radius with the largest count (vote) number to be the radius of the center of the circle (the yellow dot in our case).

In another case, say the green dot as the center, we can see r = 10 (count=5), r = 11 (count=12), r = 12 (count=15), r = 13 (count=17), r = 14 (sorry, I forgot to make this one), r = 15 (count=21). If we set the count (vote) threshold = 30, then no radius will be taken into decision. Thus, there is no any circle at the green dot.

The voting process is illustrated in the following GIF. The yellow dot is the center of various virtual circles. The red and the green dots are the examination point. If the red or the green point detects the white pixel when passing it, the vote will plus one for the current radius.

GIF 6. Voting Process ( Searching space r = 3 ~ r = 6 )

We make a little change on the vote system. When determine the radius, say at r = 13, we will also count the vote in r = 12 and r = 14 to r = 13. In this case, the vote number for r = 14 will contain the vote in r = 13 and r = 15.

The count (vote) threshold is not a fix value. It will vary with the radius. The larger the radius is (thus longer circumference), the larger the threshold should be. The calculation for the threshold is [ 2 * π * ( r - 1 + r + r + 1 ) * Accuracy / 3 ] * [ 3 / 2 ] ==> 3πr * Accuracy where the accuracy is given by the user.

If the accuracy is set too high, then the application will hardly detect any circle (because the criteria is too high to define a circle). If the accuracy is set too low, then it will be susceptable to the noise and give many circles that are definetely not circles. The following image demonstrate the relationship between the accuracy and the result of circle detection. Circles in the Image 3 have been merged. The term "Misdetect" means the algorithm thinks there is a circle, but the truth is that there is no circle.

Image 5. The relationship of the accuracy and the result of circle detection.

Merge Circle

The algorithm might draw many circles which are pretty close to the ground truth. The merge process is to determine which circles should be combined together because they are indicating the same ground truth. Circles which are close enough to each other will be assigned to the same group. We define Distance of two center of distinct circles < ( radius of one circle + radius of another circle ) / 2 as close to each other. The determine process is illustrated in the following GIF. We examine each of the circles provided by the Circle Hough Transform Algorithm. Thus, the number of groups is equal to the number of circles.

GIF 7. Define the group of circles

The circles in the same group will be combined to one circle. We use the concept of center of mass to claculate the center of merged circle ( x merged , y merged ) and the radius ( r merged ).

r merged = ( r 1 + r 2 + r 3 + ..... ) / number of circles in the group
x merged = ( x 1 * r 1 + x 2 * r 2 + x 3 * r 3 + ..... ) / ( r 1 + r 2 + r 3 + ..... )
y merged = ( y 1 * r 1 + y 2 * r 2 + y 3 * r 3 + ..... ) / ( r 1 + r 2 + r 3 + ..... )

We do the merge process four times to further combine the circles and the result does improve. Take the GIF 8 for example, Group 1, Group 2, and Group 3 will form three separate merged circles. However, they are indicating the same ground truth. Thus, we require another merging process to further combine these three circles which are so close to each other.

Some Cool Function

Make Heavy Work in The Background

For those who are new to Windows form might be surprised that their application just hang there without updating any textboxs or labels untill all the calculation are over. With the BackgroundWorker class, we can easily update the progressbar as well as other variable showing on the form and your application will not freeze when you try to click other buttons during the heavy calculation of the application. Here are two links for your reference : https://docs.microsoft.com/en-us/dotnet/api/system.componentmodel.backgroundworker?redirectedfrom=MSDN&view=netframework-4.7.2 and https://www.wpf-tutorial.com/misc/multi-threading-with-the-backgroundworker/

Devides Image into Three Parts and Process them together

Our image will be devided into three parts (two green and one blue) like the following image when the auto measurement started. Three backgroundworkers will be initialized for processing these three parts individually so that we can speed up the calculation time.The overlape (denotes as Y in the image) is to make sure those circles on the edge will be detected correctly. Each backgroundworker is able to post its own progress during the calculation (so that you can see the progress bar is updating the progress and see the circles are being detected by the algorithm in real time) and give us it's own result when complete.

Image 6. The image will be devided into three overlapped parts and processed by three backgroundworkers.

Notice that an object can only be modified by one backgroundworker at a time. It will be at risk to post each progress from three backgroundworkers to a PictureBox. Thus, we give each backgroundworker a PictureBox for updating their progress. The backgroundworker which handle the upper segment of image will post it's picture in PictureBox1 while the one which handle the center image will post to PictureBox2 (larger than PictureBox1 and PictureBox3) and of course the one which handle the lower segment of image will post to PictureBox3. These 3 PictureBoxes are resizable to fit the original image width and height ratio. PictureBox1 is always located in the upper left corner of the panel which you can refer to image 7 in the following. The location and the size of these 3 PictureBoxes are calculated in the following c# code. The progress bar is only updating by the backgroundworker which handle the image in the center (it takes more time to finish due to the larger segment of the image compare to the other two.) Not untill the last backgroundworker completes it's task, will we collect all the circles from each of the backgroundworkers to an array. This array will go through 4 merge process and then the merged circles will be drew on the original image (without edge detection) and posted on the other PictureBox.

Image 7. Those three PictureBoxes in the UI design are for three backgroundworkers to update their circle detection progress.

If the original input size of the image is (imageW,imageH), then location and the size of these three PictureBoxes are calculated as following

//original input size of the image is (imageW,imageH)
int Y = maxR; // maxR is input by user

if (Y > (imageH / 3))
{
      Y = (imageH * 1 / 3) * (2 / 3);
 }
 
int total_H = 2 * 555 * imageH / (3 * imageW) + 555 * imageH / (3 * imageW);
int total_W = 555;
int p1H = (int)(555 * imageH / (3 * imageW));
int p2H = (int)((555 * ((imageH / 3) + (2 * Y))) / imageW);
int p3H = (int)(555 * imageH / (3 * imageW));
int p2y = (int)((555 * imageH - 3 * Y * 555) / (3 * imageW));
int p3y = (int)(2 * 555 * imageH / (3 * imageW));

if (total_H > 500)
{
       total_W = 500 * imageW / imageH;       
       p1H = (int)500 / 3;
       p2H = (int)((500 / 3) + (2 * 500 * Y / imageH));
       p3H = (int)500 / 3;    
       p2y = 500 / 3 - (500 * Y / imageH);
       p3y = 1000 / 3;
       
}

pictureBox1.Location = new Point(0, 0);
pictureBox1.Size = new Size(total_W, p1H);

pictureBox2.Location = new Point(0, p2y);
pictureBox2.Size = new Size(total_W, p2H);

pictureBox3.Location = new Point(0, p3y);
pictureBox3.Size = new Size(total_W, p3H);

Zoom in to Draw Line and Measure on the Image

There is a great post which I refer to : https://stackoverflow.com/questions/51138565/how-to-measure-length-of-line-which-is-drawn-on-image-c-sharp

Notice that you should first create a panel and make the AutoSize properties to False and the AutoScroll properties to True. Second, put the PictureBox on this panel (Do not make the PictureBox size bigger than the panel). Set the SizeMode of the PictureBox to Zoom. Follow the steps in the post and then you can have the zoom in and zoom out function. Maybe the image in the PictureBox changes the position after the first zoom in or out. You can put the following code when loading the image so that the image will only change the position for this time without zooming the image ( the user will not notice that).

float zoom = 1f;
pictureBox1.Image = Image.FromFile(opbl.FileName);  // opbl is OpenFileDialog
int w = (int)(pictureBox2.Image.Width * zoom);
int h = (int)(pictureBox2.Image.Height * zoom);
pictureBox1.ClientSize = new Size(w, h);                

UI Design

I pretty like the UI design by the following youtube link.
https://www.youtube.com/watch?v=nLfzH4xOVqo

Problem Need to be Solved

The edge detection result is not ideal for some of the pictures given by our quality control department which lead to poor circle detection. Some bad image conversion result are as the following.