Home > Photography > Colour Isolation Using Filters
Colour Isolation Using Filters
This series of experiments was undertaken as the result of a discussion I had regarding the Colour Isolation article. The main point of contention was whether or not the degree to which the red, green, and blue channels of an image overlap (in terms of spectral range) could have substantial effects on the resulting image (particularly to the point that they could not be corrected by using a custom colour profile).
As you will see, this factor can have a dramatic impact on the image captured by the camera (although it is not guaranteed to), as well as the result when the red, green, and/or blue channels are used in false-colour composites along with infrared and/or ultraviolet images.
For completeness, many images are provided. If you'd like to get the high-level overview only, you can limit yourself to the False Colour Comparison and "Green Area" maps of various types, which show the extent of areas of the subject material which are "green" (this being the easiest colour to distinguish from others in all of the examples).
A pair of unfiltered halogen lights were used to provide broad-spectrum illumination of the scenes being photographed.
To collect red, green, and blue images using a variety of spectral ranges, the following filters were used:
- LDP CC1 (which I am 99% sure is an uncoated 2.5mm-thick piece of Schott BG38 - see Filters) [human-visible light]
- LDP CC1 and B+W 491 "Redhancer" didymium filter [filtered human-visible light]
- Omega 670WB35 [very deep red]
- Robert Cairns/Schott RG665 and LDP CC1 [very deep red]
- B+W 091 (Schott RG630) and LDP CC1 [deep red]
- B+W 090 (Schott OG590) and LDP CC1 [red]
- Omega 535BP50 [green]
- Omega 525DF20 [green]
- Tiffen 47 and LDP CC1 [blue]
- Tiffen 47B and LDP CC1 [deep blue/violet] 
- "Violet Stack 1.0" (Kodak 34, Tiffen 47B, and LDP CC1) [deep blue/violet]
In addition, the following UV and IR filters were used to capture additional views of the scenes:
- Baader U-Filter [ultraviolet-A]
- B+W 093 (Schott RG830) [near infrared]
- Robert Cairns/Schott RG665 [deep red plus near infrared]
- LDP BPR, BPG, and BPB [near infrared]
Using data provided by the manufacturers and independent sources (as I do not currently have access to a spectrometer), I've compiled some approximate transmission/sensitivity graphs for all of these. These graphs should not be treated as highly-accurate - they are intended simply as a guideline regarding the spectral range transmitted by each filter or filter stack. I've also included normalized versions of most of the graphs, because the difference in overall transmission is more or less removed by scaling the image levels in post-processing. For example, when normalized, the transmission of the RG665/CC1 stack is extremely similar to that of the 670WB35, to the point that I didn't actually use any of the RG665/CC1 images in the final results, because they were virtually the same except had lower contrast.
The red/green/blue combinations I settled on for the images are compared in the second set of graphs. There were many other possible combinations, but they seemed like the best overall sample of the possibilities.
Two green filters are included in the graphs but were not used in the experiment - the 11 and the 58. When I performed the experiment, I was not aware that these filters actually had restricted transmission compared to the Bayer filter in my D70 (and I do not own a green 58 filter). However, based on their curves, they appear to provide useful intermediate steps, and I will attempt to use them in a future experiment for comparison.
For more information on how this information was compiled, see the Technical Details section at the end of this article.
Test 1 - Bananas
I chose bananas for the initial experiment because I'd explicitly suggested them in the discussion that sparked it.
Somewhat surprisingly, bananas turned out to be an excellent test of my theory. As you can see, the range of "green" in the image varies greatly between the two ends of the spectrum. In the case of these bananas, it is also possible to somewhat mimic the effect in software (two examples are provided at the end of the image set).
In addition, an interesting pattern shows up in the blue, violet, and ultraviolet-A. This pattern is not noticeable in most of the colour images, but is easily discerned in the greyscale comparison. I was so surprised by this that I used it as the basis for the Your Eyes Are Terrible At Seeing Blue article.
Test 2 - Apples
Apples (at least of this specific type) do not appear to exhibit much variation when different "widths" of filter are used. There are certainly variations in these images in terms of green intensity, but the location where green changes to red does not move significantly.
Test 3 - Oranges and Lemons
As if by magic, changing the boundaries for red and green causes complex green patterns to emerge on both of these types of citrus fruit.
Test 4 - Squash
To my surprise, the squash actually showed the highest amount of variation. The delineation between "orange" and "green" is so sharp (regardless of where it occurs) that the effect cannot be duplicated in software, even though the contrast between the two can be enhanced.
Test 5 - Watermelon
This watermelon did not exhibit as much variation as some of the other examples in the red/green/blue variations, but the filters' effect on the false colour results were quite pronounced.
Conclusion and Implications
The results of this initial set of tests were significantly more noticeable than I expected. Keep in mind that even though these experiments used optical filters, the same type of result can almost certainly occur based on the type of light source in a room. For example, fluorescent lighting (particularly cheap fluorescent lighting) should have an effect very much like using one of the more narrow sets of filters.
On the other hand, it seems unlikely to me that the "functional content" of photographs will be affected if those photos are of artificial subject matter (much like IR and UV photography). If a human is painting a house (for example), they're very unlikely to go out of their way to paint a pattern that will only show up with special filtering, so that house will most likely appear more-or-less the same regardless of the filter specifics. So while I still think that this effect could be the cause of the colours "not looking right" in photos I take of e.g. green trees, most photographers probably don't have a lot to worry about.
Still, I'd love to see camera manufacturers incorporate some sort of variable-Bayer-filter mechanism that would swap out the filter array for one of two or three alternate arrays with different spectral ranges.
In addition, I've thought a lot about the implications of the significantly-changing size of the green areas in three of the five test sets. What is actually happening there is not just a gradient from green to yellow (or green to orange, et cetera), but a complex gradient of narrow-spectrum changes. That is, if we could see hundreds or thousands of primary colours instead of only three, and those primary colours were spaced at e.g. 1nm intervals along the spectrum, we would see an unimaginable rainbow in every banana, orange, or squash.
For the next iteration of experiments, I plan on marking the location at which I see the end of one colour and the beginning of another, so that I can compare this to what shows up in the test images, as well as photographing the same scene with several different cameras.
This section is provided for reference to those who are curious and/or want to reproduce my results or check my methodology - casual readers can skip it entirely.
For purposes of the graphs, my assumption is that the LDP CC1 is a 2.5mm-thick piece of Schott BG38, as I have measured its thickness using calipers and its manufacturer-reported transmission curve is closest to BG38 at that thickness.
The data for the transmission/sensitivity graphs was compiled from the following sources:
- All Schott/B+W glass/filters except for the B+W 491: numeric data tables in Schott data sheets from the Schott website, which cover the spectral range from 200nm to 5150nm. I have a high degree of confidence in this data.
- B+W 491 "Redhancer" didymium filter: graphs in B+W's catalogue appear to be identical to the graph in the datasheet for Schott S8801. That datasheet does not contain a numeric table, but the graph is higher-resolution than the one in the B+W catalogue, so I approximated numeric values from the graph. I have a medium degree of confidence in this data.
- Tiffen 47, 47B, and Kodak 34 filters: Transmission of Wratten Filters, Allie C. Peed Jr. (in PDF format). This document contains numeric tables for the filters in question from 400nm to 700nm. I have a high degree of confidence in this data. The data was compared to graphs published by Tiffen and their filters appear to have very close (if not identical) transmission to the Kodak/Wratten filters of the same number.
- Tiffen 11 and 58 filters: approximated from graphs published by Tiffen. I have a medium-low degree of confidence in this data, because the graphs were quite small.
- Omega 670WB35, 535BP50, and 525DF20: approximated from high-resolution graphs provided by Omega Filters. I have a medium degree of confidence in this data - the source is extremely reliable, but I do not trust my approximation to be highly-accurate.
- LDP BPR, BPG, and BPB: approximated from graphs published by LDP, LLC. I have a medium-low degree of confidence in this data, because the graphs were fairly low-resolution and I do not trust my approximation to be highly-accurate.
- D70 sensor response: approximated from a graph published as part of a Stanford University research project, and cross-referenced with an unsourced graph included in a forum discussion. I have a low degree of confidence in this data because there are noticeable differences between the two graphs, the Stanford graph strikes me as the victim of a recording artifact due to its spikey shape, the data is fairly low-resolution, I do not trust my approximation to be highly-accurate, and finally, because I am assuming that the stock D70 sensor response is similar to that of my modified D70 with a CC1 filter in front of the lens. But it is better than nothing :\.
- Baader U-Filter: approximated from high-resolution graphs published by Dr. Klaus Schmitt. I have a medium degree of confidence in this data - the source is extremely reliable, but I do not trust my approximation to be highly-accurate.
To accurately balance the white, black, and grey levels of the images, reference photos were taken using the same filters and exposure settings as for the subject photos. The reference materials used were:
- A ColorChecker Passport.
- A piece of bright white Teflon (PTFE), which is very bright through the entire spectral range photographed.
- A WhiBal grey card (which is not neutral into the ultraviolet, and therefore appears off-colour in variations which incorporate UV-A elements).
- Two IR long-pass filters (830nm- and 1000nm-cutoff) and a UVR Defense-Tech "Andrea-U" UV-pass filter, arranged to provide overlapping areas between the IR and UV filters, so that at least one area of the image would always be "black".
The ColorChecker Passport was used to create custom colour profiles for this specific combination of lighting, lens, and filters for the "CC1 Only" and "CC1 and B+W 491" variations using Adobe's DNG Profile Editor. A simple custom profile was also created to allow the ultraviolet photos to be correctly white-balanced. The other images did not use custom profiles.
The reference images were then composited together in the same way as planned for the other sets, and balanced using the following steps in Photoshop®:
- Create a Levels adjustment layer.
- Set the black point based on the area in the image where the UV-bandpass filter overlaps with the darker of the IR long-pass filters.
- Set the white point based on the brightest area of the Teflon bar.
- Create a second Levels adjustment layer on top of the first.
- Set the grey point based on an area where a shadow has been cast on the Teflon bar.
- Duplicate both of the adjustment layers into the file containing the fruit or vegetable instead of the reference material.
For the G-B-UVA false colour composites, the blue (~380-400nm) channel of the U-Filter image was used. For the NIR-R-G false colour composites, the greyscale B+W 093 image was used.
The original reference images and several false colour variations are presented here for comparison.
The "green area" binary (two-tone) maps were created using the following set of steps in Photoshop®. The specific numbers used are arbitrary, but were used consistently for all of the images.
- Extract greyscale versions of the red and green channels.
- Subtract the red channel from the green channel.
- Apply a Levels adjustment layer to the result, with the white point set to 26.
- Apply a Threshold adjustment layer on top, with the cutoff set to 13.
- Merge the result into a single layer.
- Delete all of the content that does not represent the subject matter (using a manually-created mask).
- Perform a Select by Colour Range operation. Select black, with a Fuzziness value of 60.
- Delete the selected areas.
- At this point, only a white map of the "green" areas should remain.
- Perform a Select All, then nudge the selection to snap it to only pixels with an alpha greater than 0.
- Fill the selected area with white, twice.
For the graduated maps, the first three steps of the same process were used, and then the result was merged.