Measuring Diameter of Calibration Pin

I want to measure the diameter of the pin but getting a bad edge detection image. Can someone help me to further improve it?

My edge detection image:

My coding:

%% Read and Binarize
load image.mat
I4 = imgaussfilt(I3);
I4 = I4 <120;

%%
I5 = edge(I4,'Sobel');
imshow(I5)

Answer 1

[Image Analyst]

Why are you doing edge detection???? I see no need for that. If the edges are about where you put them, then try thresholding and getting the area and divide by the width:

mask = grayImage < someValue;
mask = imfill(mask, 'holes');
diameter = nnz(mask) / size(mask, 2); % height = area / width.  Answer is in pixels.

It's trickier if you want the inner, instead of outer, dark thing.

[Ang Xian Jia]

Hi Image Analyst. How to choose the best threshold value? I had used your interactive tool and obtain value roughly at 120.

After using your code in the comment above, even though I had used imfill function, there are still some noise on the edges. How to filter that out? Thank you.

[Image Analyst]

You should increase the backlight to get as much contrast as you can without causing too much glare on the edges of the pin. Then just pick a threshold and use that. As long as you have good control over your lighting and camera exposure, you will be able to use a fixed threshold, which you should. (If you don't, then you should put down a gray scale step wedge transparency beside the pin and adjust the image intensity so that the steps match the reference step, which is one from a "good" image that you want to define as a reference.)

Don't worry about the noise. From what I see, it won't impact your diameter at all - maybe out in the 4th or 5th decimal place, not enough to concern you. If it does concern you, then go to higher resolution. But since the edge of the pin is somewhat fuzzy, who's to say it should be a perfectly straight line instead of ragged anyway? Maybe it really is ragged. Anyway, that noise in the measurement might be less that you'd have just by taking another snapshot after replacing your pin (your repeatability). And maybe the true edge of the pin is in between pixels so that half the time the threshold has it on the line above and half the time it's on the line below. I don't think this is a major source of measurement variability. You could do a little better by removing and replacing the pin and taking 5 measurements at each angle, then doing that for 3 angles at 60 degrees to each other. This will help with non-roundness of pin (ellipticity). If even that is not accurate enough for you (maybe these are going into a space craft???) then you'll have to use an industrial CT instrument to get a true 3-D model of the pin.

[Ang Xian Jia]

Hi Image Analyst, thanks for the reply. It helps me a lot.

So actually in image processing,even though we have done all the steps correctly(adjust lighting, good threshold, etc.), how do we know the images after we have processed are considered as 'accurate' to the original image? For example, for this pin in this case, will we ever know the edge is actually straight line or ragged?

[Image Analyst]

Do you have another, more trusted method? If not, then why do you doubt the image analysis? Why do you assume it's wrong if you don't have another more trusted method? If it's the only thing you have that gives you an answer, then why do you doubt it?

What you really need to ask is if the image analysis meets your needs. For example does it properly identify diameters that are for usable parts and unusable parts? If not, you can construct an ROC curve.

[Ang Xian Jia]

In measurement field, I think it's a headache to ask myself whether the image analysis had met my need. I think using more sample as you had mentioned is the best way to know whether I had done enough step/coding to process the pin image. Thank you so much, Image Analyst! #LongLiveImageAnalyst