- To participate in the 911Metallurgist Forums, be sure to JOIN & LOGIN
- Use Add New Topic to ask a New Question/Discussion about Geology and Geometallurgy.
- OR Select a Topic that Interests you.
- Use Add Reply = to Reply/Participate in a Topic/Discussion (most frequent).
Using Add Reply allows you to Attach Images or PDF files and provide a more complete input. - Use Add Comment = to comment on someone else’s Reply in an already active Topic/Discussion.
Texture Modelling (10 replies)
It is the next frontier in Geomet. I feel sure that, given a standardized set of core photos - all the same size, color balanced, etc, it should be possible to use a training based statistical approach to classify textures throughout the dataset. This sort of analysis is well developed in some other disciplines, surely we can import it into geology. Like many other projects, I have some hundreds of Km of core photos, data waiting to be turned into information.
I would like to know what kind of data do you use in order to develop texture models. I am highly interested in this kind of models.
To some extent what you seeking already exists. There was a project: Texture Modelling in the AMIRA Geomet project. A subproject was looking at RGB images (Ron Berry/CODES). I understand he used recognition software by Wang.
The idea of texture classification via photos is debatable. On one hand given you are taking photoes it makes sense to use recognition software. On the other hand I was (and to some extent still am) more committed to using spectral analysis for identifying minerals; and therefore saw the RGB approach as a distraction. I think from a technical viewpoint there could be additional technical issues with photoes due to variations in scale and brightness. So these issues would need to be considered.
Having said that, if I were given a project (from scratch) to classify photoes I would do the following:
Scan the photoes into a computer and convert to three channels (RGB). Convert the RGB photes into pseudo minerals. i.e. each RGB image (3 channels) would be converted into a single channel.
I would then use software I have already developed via Matlab to assess the similarity of each photo to each other and then classify into textures. It would be a very interested exercise to compare the texture method using photos with the texture method using detailed hyperspectral data.
Similarly any comparison to Wang's software would also be elucidating.
Are you thinking in grain size, liberation, mineral association type of data? I have some experience on that in case you think I can help.
I guess I better elaborate. Firstly I did alot of the foundational work for the MLA. i.e. development of software to read the MLA images and write to a database (Access).
Most of my work in terms of mineralogy was largely to do with the stereology problem. This wasn't specifically made available to commercial MLA users so I ended up working independently from mainstream. I developed numerous models at JKMRC and now largely continue and take advantage of this work independently.
So my level of work was very much aimed at analysing the images directly rather than be dependent on other service providers. Now there are two broad ways to define texture:
One is a feature-based approach (i.e. grain size, liberation, mineral association, etc.)
The other is an abstract approach using say proximity function:
The probability a pixel a distance x away from mineral m is say mineral n. The abstract approach is well-understood by authors such as Serra and Davy and other probabilists, and indeed has relationship to the variogram used by geostatisticians.
[Matheron being recognised both as an image analyst (mathematical morphology) and geostatistician - but was fundamentally a probabilitist].
Now there are variations of the abstract approach, notably by George Leigh who used wavelets (PhD thesis/JKMRC). The advantage of the abstract approach is that it provides a basis for comparing two different images to discuss there 'similarity'.
So it is 'similarity' that I mean by texture. Different images can be sorted into textural groups.
Once similarity is identified, we can use the info. in a variety of ways.
- It can be used as the basis for sample selection
- It can be used to infer texture throughout the whole orebody.
- We can provide relationships between texture and physical attributes enabling extended interpretation of testwork.
- We can relate mesoscale image to microscale images.
So I am trying to workout the current level of industry interest.
You are always bringing interesting discussion to the forum! We are always looking for patterns, regardless if in mineral processing or elsewhere.
Please allow me to bring a different point of view:
I would say texture shows how minerals correlate/associate with each other, and it can be looked at micro, meso or macro scales. What you have described pretty well as two ways to define texture, I see more like two ways to translate what we see (thru MLA or QEMSCAN) into datasets. This is the key for the geometallurgy, as we want to populate a block model with information that will enable better decision making.
Anyhow, I see texture as the most effective indicator or metallurgical performance. But infortunately we rarely have access to it, as I still see in our industry a gap of understanding of the value of automated mineralogy.
As you are aware I agree 100% with your comment:
" I still see in our industry a gap of understanding of the value of automated mineralogy. "
I have my own views about this (and explained in more detail in the Group Stereology/Mineralogy). Although I remain of the strong opinion that the data needs to be more easily accessible. Once mineralogical data is in a standard image format such as BMP, TIFF, etc. it is easy to analyse the data using advanced software languages. Personally I think Matlab is brilliant; although R is a well-used alternative.
My strong view is therefore that if mineralogical data were in a standard format, it would be possible to provide a one week course to explain how to use Matlab to estimate fundamental info (such as 2D liberation etc.) and also explain how to store the data in a database (Access).
Once Mineralogists are able to analyse the data themselves the 'industry gap' would vanish immediately. Now I need to comment further that the mineralogical data I am using for the texture modelling work is Corescan. This is hyperspectral data at about 40 um.
This is not as detailed as the conventional mineralogical data (MLA /Qemscan) so we are also very interested in the problem of meso-scale to microscale. I have confident enough to say this work will move forward now. And Corescan does save the data in a standard format.
Glad to learn about your concern on textural modelling. I have over 30 years in ore microscopic assessment and textural modelling for base metal industry in India. Conventional and Image analysis is the best practice to define the ore textures and granulometric assessment as described by Ramdhor.
I , presently engaged in process mineralogical characterization of feed, middlings, concentrates and final tails of world class lead-zinc -silver beneficiation plants in India.
Also doing the MOG (Mesh of Grind) studies, grain size analysis, photomicrographic details showing the bi, tri and multi mineral locking and its impact on liberation of economic minerals for prediction of grade and recoveries.
Also doing textural assessment of sinters quality based on textural arrangement and classifying into various categories.
Thanks you are welcome to comment on my thoughts.
I would be very interested in this work and interactions going forward. I am not a specialist researcher in this field but mainly a continuously growing practitioner in the field of geometallurgy (with its many definitions depending on who you speak to). Currently for me I am striving to assist especially my diamond mining clients with the "holy grail" output of geomet, i.e. block model based metallurgical processability of diamondiferous kimberlite ores through understanding key parameters within the each geologically unique ore type found within a deposit. For me the treatability aspects are not just related to wanted mineral liberation aspects but also host rock gangue and clay mineral release which heavily affect process design and optimisation (as found in diamond ore, e.g. kimberlite, processing). I think this discussion forum is a great platform in keeping abreast of continuous development and value adding improvement to the minerals industry. I certainly like your idea of reaching a form of standerdisation which in itself will be challenging, but let take it step by step.
Thank you for this interesting discussion. I am really interested in knowing the real application of texture modelling for minerallurgists and geometallurgists. I am now finishing my PhD on mining engineering and I have developed a methodology for the automated characterisation of intergrowth types in mineral particles. This methodology is based on digital images (provided by electronic or optical microscopy) and stereology and it is easily applicable using MATLAB. If you want to know more about my research here you have links to two papers which describe the methodoly and its applicability.
http://www.sciencedirect.com/science/article/pii/S0892687513001428
Thanks. I had a quick read. The key concept is that there are geometric probability equations that allow transformation of low dimensional data to higher dimensional data, and I briefly previously mentioned the proximity function which was a fundamental component of the work of Davy (Probability Models for liberation).
The second key concept is the use of information theory - which regrettably is not something to discuss here in much depth..
The geometric probability equations are now largely ignored by mainstream stereologists and I discuss the reasons for this in my summary.
Where you and I have strong agreement is the use of data directly via Matlab.
My PhD (1994) was largely focused on analysing texture to estimate liberation, and I also pursued this further at JKMRC (including in the context of geometallurgy). This is a clear extension of texture modelling. At one level this problem is simple if one assumes random breakage, but quite difficult when breakage is non-random (or preferential).
Once again, once image data is interrogated via Matlab or other advanced software, the possibilities are terrific.
I would be interested in hearing comments from people as to their level of interest and involvement in texture modelling. My particular interest is that I was Texture classification module leader for the AMIRA Geometallurgy project, and have receive interest in seeing whether this approach can be advanced.
And I would like feedback as to the extent members see value in this approach. I was once involved in texture modelling. Technology has changed and I am working with a group called Corescan.
We have had various discussions with mining Companies and Corescan is instigating texture modelling as a sub-project. So yes it will move forward; but how quickly I am not sure. As far as I can see texture modelling has two major functions. The first is simplification of testwork.
The second is applying a Geostats model. The first approach is basically focused on identifying similar 'textures' in an ore and then ensuring that any further detailed sampling has an appropriate cross-spectrum of different textures.
This alone is a reasonable objective. The second objective is entirely new; and means we have to totally rethink about the way Geostatistics is currently being applied. The idea is that if we can identify 'textures' however defined we can then use spatial modelling (geostats) and map the texture throughout the whole orebody.
If it can be shown that texture corresponds with lab. results such as recoveries, hardness then clearly we have now mapped the whole orebody in terms of physical attributes as well.
it is my view that a geostats (or spatial) texture model can be developed. Hopefully this will be where I will head in the future as texture modelling gains momentum. Comments appreciated.