I wish I understood noise reduction. Is there a “dummies guide to noise reduction” somewhere? For mathematicians? For programmers?

It seems to work well for music – if – you can find a section of recording where only noise is present, and it does fail for music when the noise level is high in producing high frequency whistles.

For camera noise it becomes somewhat more difficult. Ideally you want to remove the random noise (shot noise plus sensor noise) from photographs taken in dull light.

For camera noise, see for example: http://www.darktable.org/2012/12/profiling-sensor-and-photon-noise/ or https://en.wikipedia.org/wiki/Deep_Image_Prior or https://apod.nasa.gov/apod/ap191127.html. It’s possible to plot camera models on a chart of Gaussian noise vs Poisson nose, to know which camera generates least random noise.

Are the noise reduction methods for music suitable for photographs, or vice versa?

mollwollfumble said:

I wish I understood noise reduction. Is there a “dummies guide to noise reduction” somewhere? For mathematicians? For programmers?

It seems to work well for music – if – you can find a section of recording where only noise is present, and it does fail for music when the noise level is high in producing high frequency whistles.

For camera noise it becomes somewhat more difficult. Ideally you want to remove the random noise (shot noise plus sensor noise) from photographs taken in dull light.

For camera noise, see for example: http://www.darktable.org/2012/12/profiling-sensor-and-photon-noise/ or https://en.wikipedia.org/wiki/Deep_Image_Prior or https://apod.nasa.gov/apod/ap191127.html. It’s possible to plot camera models on a chart of Gaussian noise vs Poisson nose, to know which camera generates least random noise.

Are the noise reduction methods for music suitable for photographs, or vice versa?

No answers?

mollwollfumble said:

mollwollfumble said:

I wish I understood noise reduction. Is there a “dummies guide to noise reduction” somewhere? For mathematicians? For programmers?

It seems to work well for music – if – you can find a section of recording where only noise is present, and it does fail for music when the noise level is high in producing high frequency whistles.

For camera noise it becomes somewhat more difficult. Ideally you want to remove the random noise (shot noise plus sensor noise) from photographs taken in dull light.

For camera noise, see for example: http://www.darktable.org/2012/12/profiling-sensor-and-photon-noise/ or https://en.wikipedia.org/wiki/Deep_Image_Prior or https://apod.nasa.gov/apod/ap191127.html. It’s possible to plot camera models on a chart of Gaussian noise vs Poisson nose, to know which camera generates least random noise.

Are the noise reduction methods for music suitable for photographs, or vice versa?

No answers?

Well, you have a broad description of noise going on here. I presume you are talking about digital noise?

mollwollfumble said:

mollwollfumble said:

I wish I understood noise reduction. Is there a “dummies guide to noise reduction” somewhere? For mathematicians? For programmers?

It seems to work well for music – if – you can find a section of recording where only noise is present, and it does fail for music when the noise level is high in producing high frequency whistles.

For camera noise it becomes somewhat more difficult. Ideally you want to remove the random noise (shot noise plus sensor noise) from photographs taken in dull light.

For camera noise, see for example: http://www.darktable.org/2012/12/profiling-sensor-and-photon-noise/ or https://en.wikipedia.org/wiki/Deep_Image_Prior or https://apod.nasa.gov/apod/ap191127.html. It’s possible to plot camera models on a chart of Gaussian noise vs Poisson nose, to know which camera generates least random noise.

Are the noise reduction methods for music suitable for photographs, or vice versa?

No answers?

I suspect you are already the forum expert on this topic.

The Rev Dodgson said:

mollwollfumble said:

mollwollfumble said:

I wish I understood noise reduction. Is there a “dummies guide to noise reduction” somewhere? For mathematicians? For programmers?

It seems to work well for music – if – you can find a section of recording where only noise is present, and it does fail for music when the noise level is high in producing high frequency whistles.

For camera noise it becomes somewhat more difficult. Ideally you want to remove the random noise (shot noise plus sensor noise) from photographs taken in dull light.

For camera noise, see for example: http://www.darktable.org/2012/12/profiling-sensor-and-photon-noise/ or https://en.wikipedia.org/wiki/Deep_Image_Prior or https://apod.nasa.gov/apod/ap191127.html. It’s possible to plot camera models on a chart of Gaussian noise vs Poisson nose, to know which camera generates least random noise.

Are the noise reduction methods for music suitable for photographs, or vice versa?

No answers?

I suspect you are already the forum expert on this topic.

I hope not. “I know nothing … nothing” Schultz. There is a wikipedia article about both audio and visual noise, https://en.wikipedia.org/wiki/Noise_reduction but I don’t understand it.

Oh dang it, I might be the forum expert. I’ve used noise reduction on Audacity. The approach there is to start with a patch of noise-only and extract a spectrum, then cover it with a predefined spectrum and subtract, I don’t really understand the technique.

For noise that is neither video nor audio – noisy data – I use the “smoothing spline” method to remove noise. Perhaps that can be adapted to video or audio? But it definitely wouldn’t be a good idea for astronomical photographs. I always have a soft spots for wavelets, and they have been used in some noise reduction of images, I don’t know how effectively. If the data is really long and you don’t care about for example the first and last 20 points then Fourier smoothing outdoes spline smoothing.

The ideal for images would be one that takes a very long time to compute. Express each image pixel as signal + random noise, then search through all possible combinations of random noise until you find the one that maximises the smoothness of the result. But if done Bayesian, as it should be, then aaaagh!

I did a little bit of work on blind deconvolution, which as somewhat similar, for removing camera shake and blur circles from images. What I did was not a great success. And I did find that as an optimisation problem it had multiple solutions which rather invalidated the method.

> Well, you have a broad description of noise going on here. I presume you are talking about digital noise?

Yes. Noise in a digital environment, digital audio and digital images. Though the origin of the noise might be analogue, such as shot noise and amplifier noise.

mollwollfumble said:

The Rev Dodgson said:

mollwollfumble said:

No answers?

I suspect you are already the forum expert on this topic.

I hope not. “I know nothing … nothing” Schultz. There is a wikipedia article about both audio and visual noise, https://en.wikipedia.org/wiki/Noise_reduction but I don’t understand it.

Oh dang it, I might be the forum expert. I’ve used noise reduction on Audacity. The approach there is to start with a patch of noise-only and extract a spectrum, then cover it with a predefined spectrum and subtract, I don’t really understand the technique.

For noise that is neither video nor audio – noisy data – I use the “smoothing spline” method to remove noise. Perhaps that can be adapted to video or audio? But it definitely wouldn’t be a good idea for astronomical photographs. I always have a soft spots for wavelets, and they have been used in some noise reduction of images, I don’t know how effectively. If the data is really long and you don’t care about for example the first and last 20 points then Fourier smoothing outdoes spline smoothing.

The ideal for images would be one that takes a very long time to compute. Express each image pixel as signal + random noise, then search through all possible combinations of random noise until you find the one that maximises the smoothness of the result. But if done Bayesian, as it should be, then aaaagh!

I did a little bit of work on blind deconvolution, which as somewhat similar, for removing camera shake and blur circles from images. What I did was not a great success. And I did find that as an optimisation problem it had multiple solutions which rather invalidated the method.

> Well, you have a broad description of noise going on here. I presume you are talking about digital noise?

Yes. Noise in a digital environment, digital audio and digital images. Though the origin of the noise might be analogue, such as shot noise and amplifier noise.

I’d like to know more about this one. https://en.wikipedia.org/wiki/Block-matching_and_3D_filtering

“Image fragments are grouped together based on similarity, but unlike standard k-means clustering and such cluster analysis methods, the image fragments are not necessarily disjoint. This block-matching algorithm is less computationally demanding and is useful later-on in the aggregation step.”

So far so good.

“Fragments do however have the same size”

Um, what!

“RGB images can be processed much like grayscale ones. A luminance-chrominance transformation should be applied to the RGB image. The grouping is then completed on the luminance channel which contains most of the useful information and a higher SNR.”

That actually makes sense. Have you ever tried brightening a low-light noisy image – the colours are all over the place, like a rainbow. So you do not want to fix the brightness without simultaneously fixing the colour.

So, time to start reading http://www.ipol.im/pub/art/2012/l-bm3d/article_lr.pdf Looks like I may be able to understand it, fingers crossed.

This image shows original image at left and two different noise reduction techniques, middle and right. I particularly like how sharp edges are not blurred (or have jpeg-like spurious oscillations) by the noise reduction method.

mollwollfumble said:

mollwollfumble said:

The Rev Dodgson said:

I suspect you are already the forum expert on this topic.

I hope not. “I know nothing … nothing” Schultz. There is a wikipedia article about both audio and visual noise, https://en.wikipedia.org/wiki/Noise_reduction but I don’t understand it.

Oh dang it, I might be the forum expert. I’ve used noise reduction on Audacity. The approach there is to start with a patch of noise-only and extract a spectrum, then cover it with a predefined spectrum and subtract, I don’t really understand the technique.

For noise that is neither video nor audio – noisy data – I use the “smoothing spline” method to remove noise. Perhaps that can be adapted to video or audio? But it definitely wouldn’t be a good idea for astronomical photographs. I always have a soft spots for wavelets, and they have been used in some noise reduction of images, I don’t know how effectively. If the data is really long and you don’t care about for example the first and last 20 points then Fourier smoothing outdoes spline smoothing.

The ideal for images would be one that takes a very long time to compute. Express each image pixel as signal + random noise, then search through all possible combinations of random noise until you find the one that maximises the smoothness of the result. But if done Bayesian, as it should be, then aaaagh!

I did a little bit of work on blind deconvolution, which as somewhat similar, for removing camera shake and blur circles from images. What I did was not a great success. And I did find that as an optimisation problem it had multiple solutions which rather invalidated the method.

> Well, you have a broad description of noise going on here. I presume you are talking about digital noise?

Yes. Noise in a digital environment, digital audio and digital images. Though the origin of the noise might be analogue, such as shot noise and amplifier noise.

I’d like to know more about this one. https://en.wikipedia.org/wiki/Block-matching_and_3D_filtering

“Image fragments are grouped together based on similarity, but unlike standard k-means clustering and such cluster analysis methods, the image fragments are not necessarily disjoint. This block-matching algorithm is less computationally demanding and is useful later-on in the aggregation step.”

So far so good.

“Fragments do however have the same size”

Um, what!

“RGB images can be processed much like grayscale ones. A luminance-chrominance transformation should be applied to the RGB image. The grouping is then completed on the luminance channel which contains most of the useful information and a higher SNR.”

That actually makes sense. Have you ever tried brightening a low-light noisy image – the colours are all over the place, like a rainbow. So you do not want to fix the brightness without simultaneously fixing the colour.

So, time to start reading http://www.ipol.im/pub/art/2012/l-bm3d/article_lr.pdf Looks like I may be able to understand it, fingers crossed.

This image shows original image at left and two different noise reduction techniques, middle and right. I particularly like how sharp edges are not blurred (or have jpeg-like spurious oscillations) by the noise reduction method.

The photo noise reduction is quite impressive.

Although using a wider aperture in the first place would make the blinds less distracting as well.

The Rev Dodgson said:

true.

The Rev Dodgson said:

mollwollfumble said:

mollwollfumble said:

I hope not. “I know nothing … nothing” Schultz. There is a wikipedia article about both audio and visual noise, https://en.wikipedia.org/wiki/Noise_reduction but I don’t understand it.

Oh dang it, I might be the forum expert. I’ve used noise reduction on Audacity. The approach there is to start with a patch of noise-only and extract a spectrum, then cover it with a predefined spectrum and subtract, I don’t really understand the technique.

For noise that is neither video nor audio – noisy data – I use the “smoothing spline” method to remove noise. Perhaps that can be adapted to video or audio? But it definitely wouldn’t be a good idea for astronomical photographs. I always have a soft spots for wavelets, and they have been used in some noise reduction of images, I don’t know how effectively. If the data is really long and you don’t care about for example the first and last 20 points then Fourier smoothing outdoes spline smoothing.

The ideal for images would be one that takes a very long time to compute. Express each image pixel as signal + random noise, then search through all possible combinations of random noise until you find the one that maximises the smoothness of the result. But if done Bayesian, as it should be, then aaaagh!

I did a little bit of work on blind deconvolution, which as somewhat similar, for removing camera shake and blur circles from images. What I did was not a great success. And I did find that as an optimisation problem it had multiple solutions which rather invalidated the method.

> Well, you have a broad description of noise going on here. I presume you are talking about digital noise?

Yes. Noise in a digital environment, digital audio and digital images. Though the origin of the noise might be analogue, such as shot noise and amplifier noise.

I’d like to know more about this one. https://en.wikipedia.org/wiki/Block-matching_and_3D_filtering

“Image fragments are grouped together based on similarity, but unlike standard k-means clustering and such cluster analysis methods, the image fragments are not necessarily disjoint. This block-matching algorithm is less computationally demanding and is useful later-on in the aggregation step.”

So far so good.

“Fragments do however have the same size”

Um, what!

“RGB images can be processed much like grayscale ones. A luminance-chrominance transformation should be applied to the RGB image. The grouping is then completed on the luminance channel which contains most of the useful information and a higher SNR.”

That actually makes sense. Have you ever tried brightening a low-light noisy image – the colours are all over the place, like a rainbow. So you do not want to fix the brightness without simultaneously fixing the colour.

So, time to start reading http://www.ipol.im/pub/art/2012/l-bm3d/article_lr.pdf Looks like I may be able to understand it, fingers crossed.

This image shows original image at left and two different noise reduction techniques, middle and right. I particularly like how sharp edges are not blurred (or have jpeg-like spurious oscillations) by the noise reduction method.

The photo noise reduction is quite impressive.

Although using a wider aperture in the first place would make the blinds less distracting as well.

> So, time to start reading http://www.ipol.im/pub/art/2012/l-bm3d/article_lr.pdf Looks like I may be able to understand it, fingers crossed.

Have read it now, understand about 1/5 of it. Parts I don’t understand include “https://en.wikipedia.org/wiki/Wiener_filter” and “|3D isometric linear
transform”. I don’t understand whether the algorithm starts “for each pixel” or “for each 8*8 pixel reference area” or “for each 39*39 pixel search area”, the choice there would have a big influence on denoising within 20 pixels or 4 pixels of the edge. There’s no way I could write the software given the information in the paper.

I like the algorithm very much, it’s startlingly empirical. They test it by taking what is hopefully a zero noise image, add Gaussian noise by hand, and then test how well the algorithm removes Gaussian noise. So it’s not necessarily ideal for shot noise, which is common in photography and has a Poisson rather than a Gaussian distribution.

I particularly like this bit. “The interest of this weighting is that it gives a priority to homogeneous patches. Patches containing an edge will be less taken into account than homogeneous ones on the border of the edge.”

Two adaptations of the algorithm that are mentioned in passing look good, too. They are:

• speeding it up
• shape-adaptive areas