Dear RG Community,
I am currently working with thermal images of natural scenes. Throughout the scene I positioned large artificial targets that are painted black, grey and white.
I would like to detect those targets (position varies) in all of my images automatically. I won´t be able to work with a 3-layer RGB but only with a gray-scale one-layer image (example attached, there are 4 artificial, square targets in the example image). My initial approach was to use the SLIC algorithm to create superpixel areas. I am still working on that option.
I guess it should also be possible to find the artificial targets by filtering their similar, neighboring pixels? I am currently working in Python and my coding skills are advanced...but not amazing:-) I could also use R but would prefer a Python solution.
Any help would be much appreciated. Thanks for your time!
René
With Flood Fill it's possible to isolate areas of the same color, or with a certain tolerance, with approximately the same color.
https://en.wikipedia.org/wiki/Flood_fill
https://de.wikipedia.org/wiki/Floodfill
In a subsequent step you should test for standard deviations of RGB in each area for judgement of homogeneity. Another analysis could investigate the histograms of pair-wise pixel distances within each area, which identifies the particular shape.
Regards,
Joachim
If your targets have some geometric shapes like parallelograms or squares, etc. I would recommend using line Hough transform to detect lines that form this shape and localize the shape.
Another better approach if you have enough data, you can use a deep learning segmentation network.
Also, if there is a spatial relationship between targets, you can use it to correct your results.
Hi Rene,
any clustering approach (e.g. k-means) should provide an appropriate way to solve the problem. After clustering, you can roll out the approach and differentiate cluster/segments using the shape (quadratic in your case) and the size of your targets.
As suggestion for further improvements, start to use brushed aluminium targets (if this is possible in your research). The high reflectivity of these targets allows you to use the radiance of the targets as further criteria for segmentation using e.g. simple thresholding.
Other than that I agree with Amr that Deep Learning is also worthwhile to look into (e.g. semantic segmentation).
Hope this helps.
With Flood Fill it's possible to isolate areas of the same color, or with a certain tolerance, with approximately the same color.
https://en.wikipedia.org/wiki/Flood_fill
https://de.wikipedia.org/wiki/Floodfill
In a subsequent step you should test for standard deviations of RGB in each area for judgement of homogeneity. Another analysis could investigate the histograms of pair-wise pixel distances within each area, which identifies the particular shape.
Regards,
Joachim
Thanks everyone. Floodfill looks great, stumbled upon it a few days ago but wasn sure about it. Thanks for testing it Joachim Pimiskern ! I am excited to test it today!
You can try machine learning approach by declaring what is wanted and what is to be discarded
Try this:
http://ipolcore.ipol.im/demo/clientApp/demo.html?id=255
You'll obtain this with default params:
http://ipolcore.ipol.im/api/archive/staticData/blobs/f/9/f9c5f6c992a808bfe757785a6de8a4ce8be4cf8f.png
Thanks, everyone. I have been playing with various algorithms and my greatest concern is the reliability of an algorithm. Most work fine after the parameters have been tuned according to an individual image. But in some images, the target pixel values and the background pixels show a very low contrast or the background is too heterogeneous. I would like to build an algorithm that is quite reliable...and most importantly, has some kind of test in the end that reassures that all four targets in the image (see first post here) have been detected. I would appreciate hearing your thoughts:-) I am learning a lot from this here at the moment! Thanks so much...
Dear Rene,
How about using the shape, size and relative location as criteria/test?
The idea would be to fine tune your selected approach for each image iteratively in a loop. After each loop check if exactly four rectangles have been found (1st test), test if the distance between each tarp is as it was in the field (2nd test), test the orientation between each tarp (3rd test), test the size of the tarp itself (4th test) and possibly the relative radient gradient between each tarp (test 5). If any criteria is not met, initiate a further iteration by changing the setting until the criteria is met. This procedure is time-intensive but may lead you to the desired result.
By the way, with tarp I mean target. I assume you use a hyperspectral tarp target for your analyses.
Hope this helps.
Thanks so much, Ulf Mallast ! I get the idea. For this study we custom-built our calibration panels for the thermal and hyperspectral sensor we used. I assume you are based in Germany? I should visit you once I am back and shout a coffee to say "thank you":-)
You are very much welcome. Let me know if I can be of further help. Yes, I am stationed in Germany. Drop by and we can discuss your work or further ideas over a coffee/beer etc.
@Sachin Kumuar: yes, it is a segmentation problem. Watershed algorithm could handle it, and as @Amr Elsawy indicated, Hough would enhance borders.
Regards,
Joachim
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python