YCbCr

Noise generator

marcor — Mon, 22 Dec 2008 19:13:22 +0100

This filter generates noise in an image. Available noise types are

Poisson noise
Gaussian noise
Uniform noise
Laplace noise
Lorentz noise

It is possible to generate (YCbCr) luminance noise only and to apply a gamma correction.

General description

for Poisson noise: the amount of photons per pixel value.
for Gaussian noise: the standard deviation of the noise distribution.
for uniform noise: the width of the uniform distribution.
for Laplace noise: the scale parameter.
for Lorentz noise: the half width half maximum (HWHM).

For all noise types together you can select whether a gamma correction should be applied on the image such that noise is applied to linear image data. After noise application the image is retransformed to sRGB. You can also select whether or not you want to have luminance noise only and for that the image is transformed to the YCbCr colour space before adding noise.

Noise types

Here the different noise types are described. The histogram of each noise type is a real application of the filter on a neutal gray image (128).

Poisson noise

A source for Poisson noise is photon counting. When taking a picture in the real world photons (light particles) arrive at a certain rate. However photons are not correlated and thus the time between photons is not always the same. Therefore given the shutter time of your camera and a photon rate you may collect more or less photons in subsequent pictures. The number of photons you actually collect is Poisson distributed.

Calculating poisson distributed random numbers takes a lot of time, especially for bright pixels and many photons. You can decrease computation time by selecting "Approximate" which approximates the Poisson distribution with a Gaussian distribution with the correct variance. As you can see in the histogram, Poisson noise closely resebles Gaussian noise. However, there is an important difference in for dark image values!

You can see the difference to Gaussian noise in the stepwedge below.

Gaussian noise

In nature nearly everything is normally distributed. Normally distributed noise is Gaussian noise. It appears in image from thermal noise in the amplifier and other effects. An important aspect of Gaussian noise is that its variance does not depend on the intensity of the image.

Uniform noise

Uniform noise rarely occurs in nature but when digitizing a signal errors take place that are uniformly distributed. Like Gaussian noise the noise variance is independent of the image intensity.

Laplace noise

Laplace noise has a double exponential distribution. The scale parameter determines the width of the distribution.

Lorentz noise

Lorentz noise has Cauchy-Lorentz distribution which describes resonances and decays very slowly. This is a so called fat-tail distribution and causes here salt-and-pepper noise.

Comparison between noise types

In the images below the noise variance is the same for all noise types at a brightness of neutral gray (128) except for Lorentz noise where the variance is not defined.

Wavelet sharpen

marcor — Mon, 29 Sep 2008 11:02:08 +0200

The wavelet sharpen plugin enhances apparent sharpness of an image by increasing contrast in high frequency space. The amount of unsharpness of the original image can be taken into account by adjusting the sharpening radius. As an option you can choose to sharpen the luminance (YCbCr) channel of the image only.

Features

Graphical user interface
Preview image
Option to sharpen luminance channel only
Selectable sharpen radius
Lossless image conversions

Description

The wavelet decomposition of an image results in multiple images with different frequency content. When amplifying the high frequency parts the recomposed image appears to be sharper than the original one. That way the frequency which should be amplified most can also be selected and a given unsharpness in the original image can be taken into account.

This GIMP plugin allows to perform this wavelet decomposition and allows the user to adjust the amount of sharpening and the radius of unsharpness in the original image. As an option you can sharpen the luminance channel of the YCbCr converted image only which reduces colour artifacts to appear (especially in noisy images). The sharpness of the colour contrast is not critical to the human eye.

Example

This picture was taken with the Nikon D80 and the Nikkor AF 50mm 1:1.8 lens fully open. Because the image is slightly out of focus and due to the wide apperture and aliasing the image is not quite sharp.

License

The plugin is distributed under the General Public License (GPL) version 2 and newer.

Wavelet denoise

marcor — Sun, 04 May 2008 11:28:25 +0200

The wavelet denoise plugin is a tool to reduce noise in each channel of an image separately. The default colour space to do denoising is YCbCr which has the advantage that chroma noise can be reduced without affecting image details. Denoising in CIELAB (L*a*b*) or RGB is available as an option. The user interface allows colour mode and preview channel selection. The denoising threshold can be set for each colour channel independently.

Features

User interface with preview is available.
Amount of denoising for each channel is selectable.
Possibility to denoise YCbCr or CIELAB channels instead of RGB exists.
Noise profiling is done.
Conversion from RGB to YCbCr and RGB to CIELAB is lossless.
Preview is capable of displaying one single channel only if desired.
Processing time is reasonably fast (around 20 seconds for a 10 Megapixel image) and independent from the amount of denoising applied.
User interface in multiple languages is available.

Details

The wavelet denoise plugin for The GIMP reduces noise in an image using the algorithm copied and altered from the UFRaw program [Udi Fuchs] (which inherited the algorithm from dcraw [Dave Coffin]). Instead of denoising all RGB channels at once, the plugin implementation allows to denoise the RGB channels in dividually and - even more useful - to denoise the YCbCr or CIELAB channels individually. The colour model conversions are nearly lossless as the internal calculations are done in floating point numbers and rounding errors are avoided. Because the algorithm implements a wavelet filtering it runs quite fast.

Example

This is part of a picture taken with the Nikon D80 DSLR camera at ISO 1600 in RAW format. The face is lit by two candles and a colour temperature setting of 2500 K was used. The file was opened with the UFRaw plugin.

The goal is not to reduce all noise. The unpleasant part is the chroma noise which appears only when shooting the camera in RAW mode at a high ISO speed. As you can see the colour (chroma) noise is almost completely gone in the second picture and the luminance noise is slightly reduced. If you look closely at the face you will notice that the image details have not been altered much, the sharpness is almost the same.

Note that this is just the result of a uniform application of the filter on the unprocessed image. Masking the filter effect with a selection is possible which could improve the result even more.

What's new in version 0.3

The most significant changes are the introduction of noise profiling and internationalization.

Noise profiling is done.
Internationalization is possible now, several translations exist.
New slider "Softness" (which replaces the old "Detail" slider, but is completely different from it) to control the thresholding mechanism.
Introduction of the CIELAB (L*a*b*) colour space.
Reorganisation of source files and new installation routine via Makefiles.

What's new in version 0.2

There are several changes that have been made to improve the plugin.

The denoising is done on linear data now. The sRGB gamma of approximately 2. 2 is compensated.
The interface has been reorganized
A 'Detail' slider has been added which affects the thresholds dependency on noise frequency.
Tool-tips have been added.
The single channel preview can now be displayed in colour.
There is a 'Reset' button which resets the currently selected channel tempor arily (during the time the button is pressed).
Normal reset buttons have been added.
The size of the dialog is saved now after it is closed.
The default thresholds are all zero now.
The progress bar was fixed and tweaked.

Apart from that minor other changes have been made.

License

The plugin is distributed under the General Public License (GPL).