Maximise local contrast

In IM, we can increase global contrast with -level or a number of other operations, and we can increase local contrast with -unsharp. If we increase contrast by too much, we push values out of gamut, so they are clipped and we lose detail. How can we increase local contrast so it is maximised, without clipping?

The method

Subtract the local mean. Values are close to zero, so we need floating-point to avoid clipping. The process will expand the range of these values.

mlc_minblr.miff contains high-frequency data (positive and negative). If we add it back to the blur (low-frequency data, positive only), this will reconstruct the original source image.

If the input ranges from 0 to 100%, mlc_minblr.miff has a theoretical range of -100% to +100%. In practice, most values will be close to zero.

mlc_minblr.miff isn't easy to see, so we "-auto-level" to put pixels in the 0 to 100% range, just for viewing:

If we double the values in mlc_minblr.miff before adding it back to the blur, this will increase the local contrast:

From the image we see that most areas have not clipped. In those areas, we could multiply mlc_minblr.miff by a larger number before adding it back to the blur. To maximise local contrast everywhere without clipping, we will multiply by a number that varies across the image. This is most conveniently done by dividing by the reciprocal, something like this ...

... where mlc_numbers.miff contains the divider at each pixel. So now we need to calculate mlc_numbers.miff. Ideally, this would be the local peaks in mlc_minblr.miff, but this wouldn't be a smoothly changing function. mlc_minblr.miff contains positive and negative values, so we take the absolute value, and blur it. This smooths the peaks, so we multiply it by the maximum absolute value (maxA) divided by the maximum blurred absolute value (maxB). (It would be easier to "-auto-level" the blurred absolute, but outliers would squish the values too much.)

First, take the absolute and blur it, and calculate the multiplier (mult = maxA/maxB)...

Now we can divide mlc_minblr.miff by the numbers in mlc_minblr_avg2.miff to locally maximise the high frequency component.

mlc_minblr_div.miff ranges from -100% to +100%, more or less. "-auto-level" would reduce this to exactly 0 to 100%, but outliers would squish this by too much, so we divide by two and add 50% to get a nominal range of 0 to 100%, and use oogPower.bat (see Putting OOG back in the box: Power modulation).

This is the high frequency component of the maximum-contrast image. For some purposes, this is what we want. Or we can maximise further, by thresholding:

However, it is more correct to add the high frequency component back to the blur. The high frequency component, mlc_minblr_div3.png, has the null point at 50%, so we also subtract 50%.

Aside: limiting the effect

As an aside: we can limit the effect of the increased contrast. The method shown in Limit min-max imposes a severe limit so we get sharpening with no overshoot, no halos, no out-of-gamut and no clipping.

call %PICTBAT%limitMinMax ^
  %SRC_GR% ^
  mlc_maxc.miff ^
  mlc_limit.miff ^
  3x3

%IMG7%magick ^
  mlc_limit.miff ^
  -channel R -separate +channel ^
  mlc_limit.miff

%IMG7%magick ^
  %SRC% ^
  %sHDRI% ^
  -colorspace Lab ^
  -separate ^
  mlc_limit.miff ^
  -swap 0,3 ^
  +delete ^
  -combine ^
  -set colorspace Lab ^
  -colorspace sRGB ^
  mlc_maxc3_limit.miff

The script

Generally, the two sigmas should be equal. If the second is larger than the first, the contrast is reduced. If the second is smaller than the first, contrast is increased.

We can use the blend percentage from 0 to 100 to control the degree of the effect, or use a value beyond that range.

Option	Description
%1	Input image.
%2	Output file.
%3	Sigma to find the local mean. Default: `2`.
%4	Sigma to find local maxima. Default: %3.
%5	Blend percentage. Default: `100`.
%6	`Col` or `Gray`. Default: `Col`
%7	Post-process, eg `-auto-level`. Default: no post-process.

By default, the process modifies the L channel of Lab. The Gray option is slightly faster, returning just the L channel.

This image is black where there is no detail and lightest where there is most detail, so we can use it as a mask for the effect. (See also Adaptive blur and sharpen.) Working with the last example above, mlc_m3b.png:

JPEG source

For the source, we can use a JPEG image. However, the detail that is revealed may be JPEG artefacts:

Increasing grizzle

As usual, we operate on the full-size image and create small versions for the web.

As before, we can use %mlcAVG% as a mask to modulate the effect. We equalize the mask to get an even spread of the effect.

From %SKIN_SRC%, an unpublished tool finds the bounding boxes for the eyes. This gives:

The script selCoordsBB.bat makes a mask, using the bounding boxes as gradient ellipses.

Scripts

For convenience, .bat scripts are also available in a single zip file. See Zipped BAT files.

maxLocCont.bat

selCoords.bat

selCoordsBB.bat

All images on this page were created by the commands shown, using:

%IMG7%magick -version

Version: ImageMagick 7.1.0-49 Q16-HDRI x64 7a3f3f1:20220924 https://imagemagick.org
Copyright: (C) 1999 ImageMagick Studio LLC
License: https://imagemagick.org/script/license.php
Features: Cipher DPC HDRI OpenCL 
Delegates (built-in): bzlib cairo freetype gslib heic jng jp2 jpeg jxl lcms lqr lzma openexr pangocairo png ps raqm raw rsvg tiff webp xml zip zlib
Compiler: Visual Studio 2022 (193331630)

To improve internet download speeds, some images may have been automatically converted (by ImageMagick, of course) from PNG or TIFF or MIFF to JPG.

Source file for this web page is maxloccont.h1. To re-create this web page, execute "procH1 maxloccont".

This page, including the images, is my copyright. Anyone is permitted to use or adapt any of the code, scripts or images for any purpose, including commercial use.

Anyone is permitted to re-publish this page, but only for non-commercial use.

Anyone is permitted to link to this page, including for commercial use.

call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc1.png 0.5
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc2.png 1
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc3.png 2
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc4.png 3
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc5.png 10
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc6.png 20
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_sc7.png 50

call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_b1.png 0.5 . 50
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_b2.png 2 . 50
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_b3.png 2 . 150
call %PICTBAT%maxLocCont ^ %SRC% ^ mlc_b4.png 2 . -150

%IMG7%magick ^ mlc_minblr_div.miff ^ %sHDRI% ^ -evaluate Divide 2 ^ -evaluate Add 50%% ^ mlc_minblr_div2.miff
call %PICTBAT%oogPower ^ mlc_minblr_div2.miff ^ mlc_minblr_div3.png