Intention¶

This is why the problem of low-light image noise reduction is studied and has led to a variety of different noise reduction methods,. In general most of the performance evaluations for these various noise reduction methods are done on small images contaminated with noise of a known type (Gaussian, Poisson, salt and pepper, etc. Using sourcenoisemodel, targetnoisemodel, and valnoisemodel arguments, arbitrary noise models can be set for source images, target images, and validatoin images respectively. Default values are taken from the experiment in 1. Gaussian noise gaussian,minstddev,maxstddev (e.g. Gaussian,0,50) Clean target clean; Text insertion. Gaussian filter is a linear filter that can effectively suppress noise and smooth the image. Its working principle is similar to that of an average filter, and both take the average value of the pixels in the filter window as the output. Two-dimensional Gaussian kernel template or convolution kernel.

Digital cameras often have quite some noise in their pictures. Some are worse than others, but it’s there. Here I’ll show you a way to get rid of that noise using the selective Gaussian blur filter.

The basic idea behind selective Gaussian blur is that areas in the photo with contrast below a certain threshold gets blurred. You could think of it like the opposite of unsharpen mask (that sharpens everything with contrast above a certain threshold.

Step 1¶

Since the selective Gaussian blur filter doesn’t have a preview and is quite heavy to compute (proportional with blur radius), it’s easier to only work on a cropped version of you picture. Select the crop tool from the tool box, draw a selection and click “crop”.

For best results, this should a part where there’s some noise you want to remove and some detail you want to preserve.

Step 2¶

The next thing is to experiment with different combinations of the settings (“blur radius” and “maximum delta”) in the selective Gaussian blur filter.

Normally I find that both settings should be between 5 and 10, where low numbers has less effect and higher numbers more, but this is very depending on the camera. In short, maximum delta should be just high enough for the filter to kick in on the noise and blur radius should be just high enough to remove it (actually blend it with neighboring areas).

For a more in depth explanation of what the settings do see https://docs.gimp.org/2.8/en/plug-in-sel-gauss.html.

After playing with the settings a bit I liked the clean look of a blur radius of 10 and maximum delta of 10. Then I checked what those settings looked like on another crop (which is always a good idea to do), I decided that it was on the very edge, and maybe even too much. There was too much loss of detail in the hair.

An option would have been to back off a bit and use settings around blur radius 7 and maximum delta 6, but then again the only offending area was the hair. I finally decided to go with settings of 8/7.

You’ll often find yourself in these arguments with yourself over when it’s too much and when it’s too little. All I can say is that you develop a feel for it, and I’m still not finished developing that feel. In any case, it’s good to know that the first things that start to look funny are often faces and hair.

The result¶

After 8 minutes my AMD-K6 300 running Linux finished the process, and I got the following (note they are resized to 640x480 from 1280x960, you can get the larger files at the bottom of this page):

First image: Before.Last image: After (blur radius:8, maximum delta:7).

Files¶

Next: The SUSAN Noise Filtering AlgorithmUp: SUSAN Structure Preserving Noise Reduction Previous: Krita gaussian noise reduction software

SUSAN Structure Preserving Noise Reduction

Introduction and Review

``Noise' can be introduced into a digitized image in many ways,starting with the lens of the imaging hardwareand ending at the digitization of the captured image. The reduction of noisewithout degradation of the underlying image has attracted muchattention in the past. However, whilst many ``structure preserving'filters have achieved some degree of success at preserving onedimensional image structure, very few have successfully preserved twodimensional image brightness structure, such as corners and junctions.

This section describes the SUSAN noise filteringalgorithm. This uses local measurements to obtain noise reductionwhilst preserving both one and two dimensional image structure. Themethod is based on image region determination using a non-rigid (infact, spatially undetermined) region model.

Krita Gaussian Noise Reduction System

The simplest noise reduction method is the sliding mean or box filter (see [35]). Each pixel value is replaced bythe mean of its local neighbours. The Gaussian filter is similarto the box filter, except that the values of the neighbouring pixelsare given different weighting, that being defined by a spatialGaussian distribution. The Gaussian filter is probably the most widelyused noise reducing filter. It has been incorporated into the designof many algorithms such as the Canny edge detector (see[9]). The Gaussian filter is pleasing because Fourieranalysis shows that the Gaussian spatial distribution gives a Gaussianfrequency distribution, leading to a good (often described as``optimal') tradeoff between localization in the spatial and in thefrequency domains. The sharp cutoff of the box filter causes it togive a noisier output than the Gaussian (see the discussion in[9]) corresponding to a less well localized frequencyresponse.

More complicated ``optimal' statistical estimators have been derivedwhich use local means and autocorrelation functions to calculate theunderlying signal (for example Wiener filtering -- see[49]). However, these methods are usually very slow.

Linear filters such as the box and the Gaussian filters tend to createmore blurring of image details than do most non-linear filters. Anincrease in their ability to smooth noise corresponds to an increasein the blurring effect.

Many non-linear filters fall into the category of order statisticneighbour operators. This means that the local neighbours are sortedinto ascending order (according to their value) and this list isprocessed to give an estimate of the underlying image brightness. Thesimplest order statistic operator is the median (see[69]), where the central value in the ordered list is usedfor the new value of the brightness. The median filter is much betterat preserving straight edge structure than Gaussian smoothing, but ifthe edge is curved then image degradation occurs. At corners, othertwo dimensional features and thin lines the median does not performwell with regard to structure preservation. The median is good atreducing impulse noise (for example ``salt and pepper' noise) wherethe noisy pixels contain no information about their original values.

There have been several variations on the median filter, for examplethe weighted median filter (see [7]) selectivelygives the neighbouring pixels multiple entries to the ordered list,usually with the centre pixels of the neighbourhood contributing moreentries. The higher the weighting given to the central pixel, thebetter the filter is at preserving corners, but the less the smoothingeffect is. At corners, other pixels within the convex image regionincluding the centre pixel are not given any higher weighting thanthose outside, so this is an ineffective way of preserving the cornerstructure. Other methods include the contrast-dependentthresholded median (see [55]), adaptive decimatedmedian filtering (see [1]) and the midrangeestimator (see [2]), which places the brightness estimatehalfway between the maximum and minimum local brightness values.

Three similar filters, the first two of which are typical of morecomplicated order statistic operators, are the K-nearestneighbour operator (see [13]), the -trimmedmean (see [5]) and the sigma filter (see[30]). The K-nearest neighbour operator takes the mean ofthe K nearest neighbours from within the ordered list. The valuewhich K is usually set to (5 or 6 when a 3 by 3 mask is used) meansthat corners and thin lines are badly corrupted. The trimmed meanoperator takes the mean of the remaining entries in the ordered listonce the first and last entries in the list (originallycontaining N entries) have been thrown away. The sigma filter takesan average of only those neighbouring pixels whose values