I noticed when viewed at web size, image look good. But if it is viewed at a bigger size, the flaw started to come out. I was wondering whether resizing photo can reduce noise?
edutilos- said:It just makes noise less obvious, but the noise is still there.
When you have less fine detail to differentiate noise from, obviously it is not so obvious.
I noticed when viewed at web size, image look good. But if it is viewed at a bigger size, the flaw started to come out. I was wondering whether resizing photo can reduce noise?
I don't fully agree. Resizing, if is reduction, does reduce noise, should the algorithm uses is an averaging filter or if the resizing kernel consult surrounding pixels to obtain the output pixel.
Noise is mostly inaccuracies in representation. Averaging technique on larger sampling is identical to multiple attempts on the same scientific laboratory experiments and obtain the average outcome of all samples. This helps to eliminate inaccuracies in the sampling process which contains errors.
However in a 2D image, averaging is damaging to edges of the image where there is a steep change in color values, hence slightly better algorithms which are polynomial in nature such as bicubic resampling are used to approximate the changes in the changes differential and also consult a much larger 2D sampling for better approximation.
The less detail is certainly true, but it comprise of more information than if taken with lower resolution in the first place. This is also fundamentally the same theory as why larger pixels found in full frame sensors are more advantages because of the larger surface area of larger sampling area which consolidate optically into one single pixel (simplified scenario, not considering RGB array arrangement)
I don't fully agree. Resizing, if is reduction, does reduce noise, should the algorithm uses is an averaging filter or if the resizing kernel consult surrounding pixels to obtain the output pixel.
Noise is mostly inaccuracies in representation. Averaging technique on larger sampling is identical to multiple attempts on the same scientific laboratory experiments and obtain the average outcome of all samples. This helps to eliminate inaccuracies in the sampling process which contains errors.
However in a 2D image, averaging is damaging to edges of the image where there is a steep change in color values, hence slightly better algorithms which are polynomial in nature such as bicubic resampling are used to approximate the changes in the changes differential and also consult a much larger 2D sampling for better approximation.
The less detail is certainly true, but it comprise of more information than if taken with lower resolution in the first place. This is also fundamentally the same theory as why larger pixels found in full frame sensors are more advantages because of the larger surface area of larger sampling area which consolidate optically into one single pixel (simplified scenario, not considering RGB array arrangement)
I don't fully agree. Resizing, if is reduction, does reduce noise, should the algorithm uses is an averaging filter or if the resizing kernel consult surrounding pixels to obtain the output pixel.
Noise is mostly inaccuracies in representation. Averaging technique on larger sampling is identical to multiple attempts on the same scientific laboratory experiments and obtain the average outcome of all samples. This helps to eliminate inaccuracies in the sampling process which contains errors.
However in a 2D image, averaging is damaging to edges of the image where there is a steep change in color values, hence slightly better algorithms which are polynomial in nature such as bicubic resampling are used to approximate the changes in the changes differential and also consult a much larger 2D sampling for better approximation.
The less detail is certainly true, but it comprise of more information than if taken with lower resolution in the first place. This is also fundamentally the same theory as why larger pixels found in full frame sensors are more advantages because of the larger surface area of larger sampling area which consolidate optically into one single pixel (simplified scenario, not considering RGB array arrangement)
Wah.... Mati ah~ Sounds chim to the max Edu.... :'(
My brain cannot take it... AarGghh~ Can summarize?
David Kwok said:To elaborate further on what I have explained earlier. Noises in sensors are erratic and normally doesn't have a fixed pattern. This is true spatially and temporally. What spatially means is when I use 2x2 sensor pixels taking a flat surface of GRAY (assuming the GRAY is purely RGB(128,128,128)), noises exists in the capturing and hence
you get pixel (1) with values (127,128,128), pixel (2) with values (129,128,128), pixel (3) with values (128,125,128) and pixel (4) with values (128,130,128). The presented pixel values are hypothetical but highly possible. You perform a simple downsize by reducing the image by 1/2 along the width and height, meaning a 2048x2048 image becomes 1024x1024. Now the 2x2 pixels will becomes 1 pixel value right ?
Do a simple averaged math, you get the single pixel value becomes (128,128,128). It means even noise exist, it is possible to get back the original value. Is this too good to be true ? No it isn't, but will you definitely get back the original intended value ? No you wouldn't, but based on mathematical probability, you will likely get a nearer to original value than if you take at 1024x1024 RAW pixel in the first place, because it's a 1 time sample versus I have 4 samples for a 2x2 pixels and reduce into 1 pixel value. That is also true why if you go to casino and gamble, on BIG and SMALL across infinite number of times, your chances of winner is getter nearer and nearer to 50%. In realistic, you will never be 50% because you have less money and time compared to the casino, so in the end, you always LOSE if you are willing to stay long enough. (The casino example is on the presumption that LUCK is not involved and purely mathematical. Those still studying can go and verify with a mathematician in the school if you are not convinced)
The above is known as spatial noise reduction. How about temporal noise reduction. Temporal noise reduction is widely used in video post processing to reduce noise. The samples span across time instead of space. For still photography, this is possible too. I have did it before.
Placed your camera on a tripod, use M-UP mode or timer mode to take like 10 images of the same still scene using a high ISO settings like ISO3200 ? Then stack each photograph as layers in photoshop and apply mean operations across the stack of images, you will find the output image less noisy. Follow the instruction here to understand how the operation works Adobe Photoshop CS4 * Creating an image stack (Photoshop Extended)
If you are interested to learn more about noise reduction techniques, I found some good reading material found at
Noise Reduction By Image Averaging
There is no rocket science to this, you have learned in your school. The issue here is finding how to apply the technique into real life scenarios.
There is no line drawn between a scientist and an artist. The line is often drawn yourself.
Ah!!! Now THAT explains why there are some people who shoots 10 similar shots and then stack them up! Even at lowest ISO and tripod... For "best" possible IQ, right?
david, ur going into the realm of programming and digital imaging and mathematics. even with my background of programming, videography and maths, i'm having a very hard time just reading it (but yes i get your message). i don't know if its just me or what, but ur past few posts have been getting into extremely technical stuff, which can end up confusing more people than helping them.
in short. alex ortega summed up everything nicely. so for the less tech savvy (most of us), i feel that will be good enough.
when u resize smaller, u take averages. when noise gets averaged, it can become less significant. smaller images hence can result in less "visible" noise. but on the down side, fine detail is also lost, due to averages taken.
does this make sense? let me know if i'm mistaken.