Magenta edge tinging at high ISO and a possible solution...

Started Aug 17, 2009 | Discussions thread
GordonBGood Veteran Member • Posts: 6,308
Re: PROGRESS: Version has one more tiny fix...

I completed PEF support for version and revisited the correction algorithm as per your suggestions. In fact, I turned off everything except for the "dumb" averaging with a sliding window "blur" averaging that has an effect very similar to a Gaussian blur of radius 5 pixels or less. This is clearly not really enough smoothing, as Horizontal Pattern Noise (HPN) is being added to the image, but I wanted to match what you have done without any risk of averaging out the peaks/ridges. The main difference between what I do and what you do is that instead of subtracting the resulting averages from the rows and thus making "blacks blacker than black" and also distorting the colour White Balance (WB), I compute the overall average correction so that the net result for each colour channel is zero, with the darker tones that should be black getting brighter and the brighter tones that should be black getting blacker. In this way, I don't increase image contrast or need to smudge the WB to compensate for the camera computed WB based on uncorrected values.

And the result is: that your magenta ridge is still there! So I through both the before and after correction images into ACR in Photoshop to see what is going on, with the following observations:

1) That the ridge is something like 100 pixels wide, so any of the averaging I have done wouldn't really affect getting rid of it, although I will likely reduce the averaging window size a little just to optimize the program results.

2) That there are increased amounts of hottish pixels across the ridge as compared to non-ridge pixels, but that these do not significantly affect the correction levels when comparing an averaging method which does not reject the hot pixels vs. my previous method which does not reject them. One observation with the Samsung CMOS sensor as compared to other sensors is that there seem to be a higher percentage of "outliers" in the histogram toward the positive size indicating quite a lot of "hottish" photosites, even for this exposure time of 1/125 second, but these aren't what are affecting these results.

3) That this ridge is of variable strength across the width of the sensor, with a stronger effect toward the center of the sensor. I have a possible explanation for this after analyzing you submitted results below:

dlacoutu wrote:

Well, I've carefully made the test again, and this time I have more similar results between the two methods... The dumb method has in fact a faint ghost of the ridge, slightly more visible using your application (guess it comes down to the radius used)...

Okay, here is the original pic:

The "application" pic:

The "dumb" pic:

Now, it looked to you that the last image is less magenta in the ridge area than the results of the correcter program, but what you are actually seeing is more the shift in WB towards green due to the way you do the correction and (slightly) the artificial increase in overall image contrast.

Possible reason: This effect that this correcter program compensates is very likely some sort of thermal effect over the area of the sensor that is undetected and uncompensated by the black level compensation algorithm for the vertical direction in landscape orientation. Therefore, it stands to reason that the "magenta ridge" for your sensor (my samples from other K20D's do not show this) has an extra thermal effect in this area. One possible explanation is something that has been described for other sensors as "amp glow" at highly amplified/pushed exposures as these are, which may also be spot thermal gradients over the sensor area, usually along the edges. I believe that your camera exhibits a type of "amp glow" toward the center near the bottom edge which may have the explanation that the sensor may not be properly coupled to its heat sink in this area so that the heat produced by scanning the sensor is not being dissipated as effectively as for other areas; it may be as simple as insufficient heat conducting paste in that area or the sensor not quite being flat to the heat conducting surface. This being a thermal problem is supported by the fact that there are more "outliers" in this area for the histogram plots.

Conclusion: That your sensor is somewhat a unique case and I will not be able to get rid of the ridge.

Algorithm tuning: I will keep parts of the new algorithm that try to compensate for the presence of hottish photosites due to the extra "outliers" and will tune the width of the averaging window so that it is no wider than necessary to minimize HPN before I release version likely in a few hours.

Regards, GordonBGood

Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow