Canon 5D Mark II: Banding... c'mon, still?

[rishij]

<p>Thanks Felix... exactly my point.</p>

<p>Anyway, for those of you who think I was posting 'unreasonable' examples, note that this image where I showed banding by adding some Fill Light:<br>

<img src="http://staff.washington.edu/rjsanyal/Photography/BandingInShadows.jpg" alt="" width="682" height="508" /></p>

<p>... comes from the lower left corner of this image:<br>

<img src="http://img.photographyblog.com/reviews/canon_eos_5d_mark_ii/sample_images/canon_eos_5d_mark_ii_31.jpg" alt="" width="600" /></p>

<p>In such scenes encompassing a wide dynamic range, unless you're shooting negative film, you tend to expose for the highlights to retain their color. In post-processing, perhaps I want to bring some information out of the shadows. But <strong>BAM</strong> I get struck by banding.</p>

<p>Not pleasant.<br>

-Rishi</p>

[stephen_asprey2]

<p>Yeah, well thats digital for you.<br>

Me? I'm staying with film for a while yet on my fine art photos and scanning the negs at 4000dpi.</p>

[rishij]

<p>Stephen,</p>

<p>Yeah, the ironic thing about me saying that you made a fanboy outta me is that, well, uh, I think I <em>am</em> gonna remain a fanboy of Velvia for now :)</p>

<p>Although, I could just get a Fuji S5 and expose for the shadows.</p>

<p>Cheers,<br>

Rishi</p>

[rishij]

<p>Berg,</p>

<p>You had previously linked <a href="http://www.pbase.com/terrylovejoy/image/50670364">here</a> to show how to get rid of fixed column (row?) amp noise.</p>

<p>Looks nice, but how do you go about doing it? Does Canon/RAW converters attempt this themselves? Or is it something we can implement by analyzing time-invariant noise from dark (black) exposures?</p>

<p>Rishi</p>

[hawkman]

<p>Brag,<br>

you can certainly grow oxide by epi, it is going to be single crystal oxide which is basically glass, It has even been done with MBE. microlense in not organic or photoresist, photoresist is light sensitive and will change over time.<br>

Carriers in a photo diode are generated optically and not by doping, Freezout is not an issue in ANY CMOS device, 1E18 is low doping typical N-doping today is in the order of 1E20/cm^3. Even with a 5E14 doping conductivity is much better at 77K than 300K, it is simple physics. At 300K with Nd=5E14 you get 4.93E14 electrons/cm^3 while at 77K you get 3.11E14 electrons/cm^3 while mobility is increased by 2 order of magnitude, just look at a smiple electrical conductivity vs temprature plot for bulk silicon in any basic semiconductor book.<br>

The dominating factor in any detector, solar cells etc. is bulk SRH recombinatio at the junction which improves at low T.  Any ways.</p>

[scott_ferris]

<p>Arash, your technical knowledge is beyond my questioning, sometimes, but it staggers me that somebody that researches this field in the USA can't spell temperature, ever!</p>

<p>As I said in the other post with the same question, the worst camera I have used for this is the 1D, (it has a CCD sensor) but it is very easy to allow for, the exposure latitude is comparable to Velvia, anybody used to exposing for slide film would never have an issue with this stuff.</p>

<p>Anyway back over to you academics, I have to go do some more processing.</p>

<p>Take care, Scott.</p>

[hawkman]

<p>Brag,<br>

FYI, here is an abstract (public) from U.S. Patent 6001540 about how microlens array is fabricated<br>

"<strong>A process is described for forming a microlens, either directly on a substrate or as part of a process to manufacture an optical imaging array. The process starts with the deposition of a layer of silicon oxide over the substrate, said layer being the determinant of the lens to substrate distance. This is followed by layers of polysilicon and silicon nitride. The latter is patterned to form a mask which protects the poly, except for a small circular opening, during its oxidation (under the same conditions as used for LOCOS). The oxide body that is formed is lens shaped, extending above the poly surface by about the same amount as below it, and just contacting the oxide layer. After the silicon nitride and all poly have been removed, the result is a biconvex microlens. In a second embodiment, a coating of SOG is provided that has a thickness equal to half the microlens thickness, thereby converting the latter to a plano-convex lens</strong>. "<br>

Also interesting enough, I found an online process for making microlens array by etching oxide:http://www.mems-exchange.org/catalog/P3433/<br>

I think the organic material you are refering to is used in the CFA (color filter array), which is a seperate layer on top of microlens array.</p>

<p> </p><div></div>

[hawkman]

<p>Scott,<br>

I see I dropped an "e" from temperature-I am sorry-I sent that message from my blackberry with its tiny keyboard, I apologize for the typo, you have to excuse people for typos on internet forums. Also, I know very little about many things, CMOS devices and low tempErature characterization just happen to be my area of research. I look forward to learning more about photography from all the experts on this web site.<br>

Thank you for pointing this out.</p>

[rishij]

<p>Who's 'Brag' anyway?</p>

<p>And while I love academic talk, I'm still wondering if there's any practical way to get rid of the banding in software... Berg had given me hope earlier that there is. Can anyone else chip in as to the actual process? What software? What algorithm... etc.?</p>

<p>Rishi</p>

[rishij]

<p>Arash,</p>

<p>You're right that the banding is non-evident in the ISO 50 image -- because ISO 50 is in fact ISO 100 overexposed by one stop... which essentially means that there were not deep shadows in the exposure. Hence no banding.</p>

<p>So, since you're an expert on CMOS... is it just a quality control issue that some cameras show this defect more than others?</p>

<p>Is the only way, in the end, to reduce this problem by using lower gains on amps and making higher sensitivity photocells? If I understand correctly, they lowered the amp gains in the beginning when introducing CMOS to these cameras in order to reduce non-uniformity in dark areas.</p>

<p>Rishi</p>

[steve_hovland]

<p>Not seeing this on Nikon.</p>

[hawkman]

<p>"So, since you're an expert on CMOS... is it just a quality control issue that some cameras show this defect

more than others?"<br>

 

<br>

Dear Rishi, I am not an expert in commercial image sensors and I don't work for Canon, I just know

a little b

it...<br>

Rishi, all cameras have banding because there is always mismatch between the sense amplifiers whic

h will be evident when you amplify such a small signal from darkness, even if there is no mismatch and you use unity

gain the signal level will still be slightly different because you are operating close to the noise floor of the sense

amps. The only way you can take better pictures in darkness is larger or more efficient pixels to raise the sig

nal level.<br>

Some of the banding also comes from cross talk and mismatch between signal pathways that

connect the sense amplifiers to buffers and ADC. I know that in DRAM for example this issue is corrected by equaliz

ing signal paths by using a folded structure, however this is not doable in an image sensor. I am sure the

re are many other circuit tricks I am unaware of. I think Sony/Nikon use a column parallel design which places ADCs

at each column and thus shortens the

analog pathway which might result in lower banding. You can also average for longer like your scanner but then you wi

ll have a very slow camera.<br>

Like I said I don't see banding an issue with any of the cameras on the market

today at least up to ISO 3200 (the max I use) At 12,800 or 25,600 images from all these cameras are plain noisy a

nd useless with or without banding. At low ISOs I haven't seen much banding from 5DII files yet, can you post a raw

file with shadows that you think might exhibit significant banding? here is a crop form a 5DII image po

sted in<div></div>

[hawkman]

<p>hum looks like photo.net truncated my post! but I was trying to say that the above crop at ISO 3200 pushed by 2.0EV shows a lot of noise but no banding.</p>

[carnagex_carnagex]

<p>Banding on the 5D and probably the 5D MkII has to do with the lenses used on it. The 50mm 1.4 motor caused banding on the 5D. (Along with a few micro-ultrasonic lenses).</p>

[rishij]

<p>Arash,</p>

<p>Here's a link to an ISO 100 RAW image from photographyblog.com that exhibits it (same image I posted above). Increase the exposure or brighten the shadows, and you'll see the banding in the building/windows.<br>

<a href="http://img.photographyblog.com/reviews/canon_eos_5d_mark_ii/sample_images/canon_eos_5d_mark_ii_07.cr2">http://img.photographyblog.com/reviews/canon_eos_5d_mark_ii/sample_images/canon_eos_5d_mark_ii_07.cr2</a></p>

<p>Carnagex: That sounds like a rather implausible explanation to me, but, it'd be easy to test: just take some images of black with the lens off (assuming that's possible!).</p>

<p>-Rishi</p>

[s._lee_cashman]

<p>From what I've seen thus far, banding starts in at around 1600 and gets ridiculous by 6400 (IQ = barely usable, if at all). </p>

<p> </p>

[berg_na]

<p>Arash - I can tell that you are not very familiar with image sensor technology. Microlenses are formed on top of the color filter array, not under it. Glass is simply not a crystalline material. Another practical problem: The various alumimum layers used for interconnects in a CMOS chip will melt at the temperatures required to clean the wafer surface prior to epitaxial growth...<br /><br />Rishi - Sorry for the distraction... The banding you're seeing is not the fixed pattern noise caused by slight threshold offsets between the column readout amplifiers discussed in the reference I pointed you to. Fixed pattern noise correction is simple and automatically performed in the camera FPGA</p>

[rishij]

<p>Berg -- if it's not fixed pattern noise, then why does Igor see this when layering 16 different Canon 20D exposures and averaging (i.e. with one exposure it wasn't evident, but averaging 16, it became evident):</p>

<p><a href="

<p>Igor, are you here yet?</p>

<p>Thanks,<br /> Rishi</p>

[hawkman]

<p>Berg,<br />Glass in your window is amorphous, glass above refers to quartz which is <a title="R hombohedral crystal system" href="../wiki/Rhombohedral_crystal_system">rhombohedral </a>single crystal: <a href="http://en.wikipedia.org/wiki/Quartz">http://en.wikipedia.org/wiki /Quartz</a> <br />Microlens is aligned to the CCD/CMOS chip post processing so the thermal budget is not altered. Also CMOS sensors (like any other device) use Cu for interconnect no Al, Al has not been used in an y major CMOS device since early 2003, one reference is here <a href="http://www.physorg.com/news5184.html">http://www.ph ysorg.com/news5184.html</a> <br />Any ways, since you are the expert in image sensor technology who is a humble electrical engineer and a physicist to disagree. I go back to my work but you may want to read this excellent article here <a href="http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/">http:// theor y.uchicago.edu/~ejm/pi x/20d/tests/noise/</a><br>

banding noise:<br />"In terms of its spatial variation, read noise is not quite white. Upon closer inspection, there are <strong>one-dimensional patterns</strong> in the fluctuations in Figure 2. Because the human eye is adapted to perceive patterns, this <em><strong>pattern</strong></em> or <em><strong>banding noise</strong></em> can be visually more apparent than white noise, even if it comprises a smaller contribution to the overal l noise. Pattern noise is more problematic in older models such as the Canon 20D; see figure 3. "<br>

quoted from open source article linked above.<br>

<br />And because you are an expert I am sure you have access to IEEE database, so I would also recommend reading this paper from one of my great professorrs, Abbas El Gamal, who is the undisputed authority in image sensor technology not only here at Stanford but around the world.<br />I quote fro technical paper</p>

<p align="left">"In addition to temporal noise, image sensors also suffer from FPN, which is the pixel-to-pixel output variation under uniform illumination due to <strong>device and interconnect mismatches</strong> across the image sensor array. These variations cause two types of FPN: offset FPN, which is independent of pix el signal, and gain FPN or photo response nonuniformity (PRNU), which inc reases with signal level. Offset FPN is fixed In addition to tempor al noise, image sensors also suffer from FPN, which is the pixel-to-pixel out put variation under uniform illumination <strong>due to device and interconnect mismatches</strong> across the image sensor array. These variations cause two types of FPN: offset FPN, which is independent of pixel signal, and gain FPN or photo response nonuniformity (PRNU), which increases with signal level. Offset F PN is fixedz from frame to frame but varies fro m one sensor array to another. Again, there are more sources of FPN in CMOS image sensor s than CCDs introduced by the active readout circuits. The most serious addi tional source of FPN is the column FPN introduced by the column amplifiers . Such FPN cancause visually <strong>objectionable <em>streaks</em> in the image</strong>"<br>

 

<p>Source : <em>CMOS image sensors,E l Gamal, A. Eltoukhy H.Stanford University,IEEE circuits and Devices Journal, June 2005, issue 3 pages 6-20<br /></em><br />Quotation from technical journal with author permission.</p>

 

<br>

I am sorry but this issue is closed.</p>

[igor_jarm]

<p>Rishi, I'm here.<br>

Just wanted to say, no need to repeat the test without the lens, it was tested with an EF 50 1.8 II, so no USM.<br>

It might be interseting to repeat the test thou, just to see if fixed pattern noise changes with time. The test was done over a year ago.<br>

At least on 20d in-camera jpeg conversion is realy bad and the banding becomes evident even in relatively bright areas shot at ISO 800.<br>

For me the problem was removed almost completely just by shooting raw and converting to jpeg out of camera. For example see this <a href=" at equivalent of ISO 25600</a> .</p>

 

[hawkman]

<p>Rishi,<br>

Please see the references above on banding noise, or streaking artifacts. You should not trust every claim you see in an internet forum without valid citation. also note that temporal fluctuation and random cross talk might cause streaks to appear at different locations in successive images.<br>

Going back to the RAW file, I converted it using DPP 3.5.1 with +2.0EV EC. you are correct, there is banding artifacts in the shadows more than I expected, it even shows at 50%. I searched through my 40D images and I found two extremely underexposed samples one at ISO 100 and another at 400 to see if I can reproduce banding by +2.0EV EC. To my surprise 40D RAWs did not show banding as much as 5DII (see the examples below). It might be possible that something was wrong with the particular camera you were testing, or if live view had been used extensively it might have raised the internal temperature of the components and aggravating this issue, although this is very unlikely. But I will keep an eye on more samples, the rest of the samples I had seen so far did not exhibit excessive banding.</p>

<div></div>

[hawkman]

<p>40D</p><div></div>

[hawkman]

<p>crop</p>

[hawkman]

<p>here</p><div></div>

[rishij]

<p>Hmm... the crop seems to be missing?</p>

<p>Also, this could be an inter-unit variation, and not specifically 40D vs. 5D MII, right? Though probably more a sensor-based issue, so inter-model variation.</p>

<p>I don't know. I wasn't testing this camera... these were taken from photographyblog.com, just so you know :)</p>

<p>By the way, your crop of the 5D MII image shows it even better than mine... thanks Arash :)</p>

<p>Thanks,<br>

Rishi</p>