Overlapping Particles Particle Size Analysis


This image is some dust taken from an aircraft engine test and I need to analyse the particle size distribution. There are a few different minerals in here, but hopefully should somewhat be able to represent the atmospheric dust.

I am extremely new to image j and have barely learned the ropes yet.


As you can probably tell, the image is made extremely noisy by the fact that there are a lot of very small particles overlapping some of the larger and vice versa.

I have tried watershedding and modifying the threshold but they are not able to solve the image.

I have viewed the below forum page which has linked to some clever sounding stuff written by Thorsten Wagner, but I don’t know how to make use of it/whether it is applicable in my case or not. Would someone please be able to help?

Hi, i need help on separating overlapping starch particles - Image Analysis - Image.sc Forum

Interesting task …

Your sample image is unsuited for quantitative analyses mainly due to heavy over-exposure. Please check its histogram.

Furthermore, please work with and provide uncompressed images.
(JPG-compression is lossy and introduces artifacts that can’t be removed.)

Did you consider Fourier power-spectral analyses?


Hi notQRV!

I have attached the histogram here. Please again accept my apologies, I am very new to this and don’t know much about image processing at all!

I havent looked into Fourier power-spectral analyses, would that be something worth doing? What is its benefit exactly?

The gray-value histogram I get from the square-sized area (649x649) at x=90, y=0 of your sample image is:
Histogram of image
It clearly shows that the image is saturated, i.e. the highlights are clipped (value 255) and can’t be distiguished.

Regarding power-spectral analysis, it appears rather difficult to explain it to someone not familiar with signal theory and analysis (here pictorial signals). The topic is far from basic.

In short:
The Fourier power-spectrum reflects global properties of an image. These properties are essentially measures of structural coarseness.

Either you try to find an introductory text or a colleague that/who can help you with this topic.

Hi notQRV.

Thank you for your input again. I will look up Fourier power-spectral analyses and see what I can come up with.

Would you say this image is hopeless for quantitative analysis?

Thanks to your help on a seperate thread, I have been able to install the particlesizer plugin and produce the following image:

Clearly this result isn’t perfect, and the settings aren’t optimised either for the type of image I have here.

Let me know what you think if you get a moment.

It very much depends on what kind of analysis your are applying.
It don’t know the ParticleAnalyzer-plugin but if it gives you reasonable size estimates, why not use it. The shown particle segmentation doesn’t really convince me. But the decision and perhaps optimization is up to you.

I have no idea about possible output formats of the SEM-device but if it is able to deliver 16bit TIF images, you should use this format. By doing so you may avoid the gray-value clipping (over-exposure).

It is a pity that this interesting topics are discussed in the background. What is the reason?
Think I have seen this style in the past.

Hi Peter, the link to the thread mentioned is here

1 Like

Not sure, but the ParticleAnalyzer-plugin appears being “Analyze Particles…” that is part of plain ImageJ. Consequently, there is no need to install anything.

If this is true, I don’t think it will be of great help for getting a reasonable distribution of the actual particle sizes. The problem with “Analyze Particles…” is that it evaluates binary-valued image only and thresholding of your sample images will not result in representative binary images, i.e. their distribution of the particle sizes will most likely differ considerably from the actual one. Especially small particles on top of the bigger ones and those in the dark will be missing, etc.

Hi notQRV,

It appears to me (as a genuine newbie) that the ParticleSizer is much better at distinguishing the particles in my image. A snapshot from a quick attempt using the “analyse particles…” function with auto threshold is below. Not to mention the much more extensive measurements and post processing from the ParticleSizer

The poor quality of the image is clearly a huge factor here, but the ParticleSizer uses local thresholding automatically which I guess is the reason it is so much more effective in my case?

Could you please post a link to the “ParticleAnalyzer” you use so that we can check what you are dealing with.

Apart from this question, I still doubt that you will get a reasonable distribution of the actual particle sizes from these images. What you need is some kind of reference image or data with known properties so that you can quantitatively compare the results of your attempts. Everything else is pure conjecture and far from good science.

Apologies, i switched from particlesizer to particleanalyzer…

I have corrected above…

The plugin I’m using is linked here

1 Like

OK, thanks for clarifying the issue.
(It is not a good idea to change posts in this fashion. It breaks the flow of arguments …)

If you look at the sample images given at the ParticleSizer website, you may realize that there, the particles appear as essentially flat. This is not the case with your images and this property makes it really difficult to distinguish e.g. between a slanted surface of a bigger particle and a smaller isolated particle.

To obtain reasonable distributions of the actual particle sizes from your images will require advanced methods such as shape/texture based classification and I’m not sure if such will suffice.

1 Like

No problem, and thank you for helping!

I have only been in my PhD a month or so and I’m an engineer by trade so much of this is very new to me.

The topography (surface roughness) of the sample has been cited by some of my peers as the main issue with the analysis of these samples we have been given and I agree with you that the image I am attempting to analyse here is very different than those quoted in the ParticleSizer website.

As you also mentioned earlier, the extremely blown out highlights in the image also (I assume) make it harder for any computer software to be able to easily determine a good threshold for the image.

I am currently in the process of comparing the data from the ParticleSizer to that of manual methods (mean linear intercept and a manual grid technique). I will let you know how the comparison turns out.

1 Like

Traditionally signal/image processing is an engineering disciplin (mainly taught in electrical engineering) but I know of no university engineering study in which for instance the Fourier-calculus is not taught.

In fact, signal/image processing is mainly mathematics and mathematics is also the basis of any engineering.

This certainly is a good approach and you should consider doing the whole job manually. The earlier you start …
I think about 1000 of such images can be managed in about a month.

You should also re-consider the task as such, i.e. what is the desired outcome. Is it really absolute size distributions, or is it relative size distributions, i.e. changes?
In the later case, coarseness measures may be sufficient.

Perhaps the most important aspects when evaluating a PhD-thesis is originality and independence of work. Keep this in mind.

Yes I was certainly taught it in my first year of university 7 years ago however have not had the need to apply/relearn it since, though this might seem like a good time to try.

Absolute size is definitely relevant, since it is likely the results of these analyses are going to be inputted into some fluid dynamics calcs, to assess the likelihood of particle/surface interactions.

Absolutely right on that last point. This discussion is what is required as part of critical reading/technique evaluation. Thank you so much again notQRV.

To be able to help a bit further, it would be useful to see sample images that show clearly different size distributions as well as images that should be distinguished but show only slightly different size distributions.

From my point of view the two sample images that are presently available appear to belong to the second category.


I have done some more digging with some quite interesting results and I thought I would share them for maybe someone will benefit.

Here I have plotted the results using the particlesizer on a raw image comparing it to using the particlesizer on some tracings i did and a manual measurement of the tracing.

I think most would agree that the results are quite good in favour of the particlesizer on the raw image. Given the poor sample prep, I would say that its actually quite remarkable.