# How can I measure noise from the photo?

With higher ISO comes "bigger" noise. Is there a way to measure how much noise does the photo have? Maybe some programs can calculate it?

• Film has grain. Faster film has larger grain. Digital images have noise. Images with more noise are just that, images with more noise. But the noise is not necessarily "bigger", there is just more of it spread throughout the image. Commented Mar 5, 2016 at 14:24
• Thank you for the answer. Is there a way to measure noise? Commented Mar 5, 2016 at 14:27
• There are basically three reasons I can think of to do this. First, to compare cameras to each other. Second, to compare the same camera in different situations so you are later more comfortable in using the camera in those situations. And third, as part of assigning some sort of goodness score to a photograph without actually looking at at. Which of these are you interested in? Or, something else entirely? Commented Mar 5, 2016 at 14:49
• I did, thank you for the reminder. I'm doing my Bachelor's degree in Photography and I'm comparing two photos from different cameras with the same ISO. How could I measure which one has more noise? Commented Mar 5, 2016 at 14:59
• Are they two photos of the same scene in the same conditions? Are you comparing for (literally!) academic interest, to learn about noise, or for some other reason? Commented Mar 5, 2016 at 17:23

Noise is the difference between what the sensor measures and what it should have measured. If you take a photo with the lens cap on (or with some other means of completely blocking any light from entering the lens), you'd expect to get a frame where every pixel is entirely black. In reality, you'll get an image where the pixels vary slightly from one to the next due to noise in the sensor. Try it with different ISO settings and you'll find that higher ISO settings give you more variation between pixels in a dark frame.

Given that, you could calculate the average brightness of the pixels in a dark frame and use that as a simple measure of noise. Comparing those averages for different ISO values would give you an idea of the relative level of noise at different settings. It's far from perfect, though -- noise can presumably be positive or negative, and by pegging the expected value at zero we're basically ignoring half the effect of noise.

The important idea here is that you know what the value for a given pixel should be. Using a dark frame isn't the only way to know that. Another way is to take a photo with the lens intentionally defocussed, e.g. take a photo of a distant subject with the lens focused at its closest setting, or vice versa. This should give you an image where changes from one color to the next should be very gradual -- differences between adjacent pixels should be very small, and any large differences between one pixel and its neighbors are probably attributable to noise. You'll probably want to get a little more sophisticated in your analysis here, maybe using the standard deviation instead of just a simple mean.

Even without defocusing, you can measure noise in an image by looking for areas in an image where you have smooth transitions between colors, such as in the sky. Noise generally affects the entire sensor in more or less the same way, so it's reasonable to measure the noise in just part of the image and assume that the noise level will be similar across the entire image.

There's lots more to know about measuring noise in images. If you really want to do it right, you should read about signal to noise ratio (SNR) and how to calculate noise levels in decibels (dB). Searching with terms like image noise measurement software will help you find existing tools that can help you.

• I think you may be surprised about the noise distribution. It's frequently related to heat so it's likely you'll find hotspots. That's why the astronomy folks frequently cool their cameras and some sell mod kits (like this one for the Canon 700D which cools the sensor to around -15 to -20°C! Commented Mar 5, 2016 at 21:00
• @JamesSnell Read noise is subject to hot spots and heat. Photon shot noise is random and the distribution of shot noise will not be affected by warmer areas of the sensor. Commented Mar 5, 2016 at 22:09
• @MichaelClark Let me just follow up on that a little bit. Shot noise also does not affect the entire sensor the same. It grows with square root of light intensity, and so does its SNR as well (x/sqrt(x)=sqrt(x)). So pixels in the image with four times the intensity get double the shot noise and double the SNR. Commented Jul 25, 2018 at 18:05
• @relatively_random Shot noise is randomly distributed, but only on a microscopic scale. Assuming the field intensity of light is the same within the lens' FoV, shot noise will be more or less evenly distributed over the entire sensor when measured in areas that cover, say, 1/10,000 the surface of an APS-C sensor instead of 1/20,000,000 the surface of the same sensor. Commented Jul 26, 2018 at 2:39

If you're interested in measuring camera noise, Dr. Roger Clark has a number of articles on the topic, including on how to measure it.

Measuring noise in a non-controlled image is a bit more difficult, though. You could try estimating the noise levels from the smooth areas. To do that, you could try something along these lines:

• First convert the image to a format which can handle negative numbers (floating point or signed integer).
• Heavily smooth the image (gaussian or median filter, for instance).
• Subtract the smooth image from the original to get the difference image. In uniform areas, this should be just noise: some values fluctuating +- around zero.
• For each of those areas, calculate the standard deviation of the values to estimate noise level.
• Divide the average value of the smoothed image in the selected area by the standard deviation to estimate signal to noise ratio.

Keep in mind, these numbers will describe per-pixel noise. A higher resolution camera of the same format and the "same" sensor quality will inherently be noisier when compared like this. Per-pixel noise doesn't tell you how noisy the object appears to be in the image, just how noisy a small part of it will look when enlarged so you can see the pixels.