Image Noise in Digital Cameras

To some extent, the pop­u­lar demand for increas­ing­ly high­er ISO per­for­mance is an expres­sion of pho­tog­ra­phers’ desire for high­er speed. Here, the word alludes to the same con­cept referred to by lens speed, relat­ing to the shut­ter. When dig­i­tal cam­eras became wide­ly avail­able, the flex­i­bil­i­ty of vari­able ISO val­ues offered dig­i­tal pho­tog­ra­phers a unique glimpse into the world of pos­si­bil­i­ties for low-light hand­held pho­tog­ra­phy that would oth­er­wise require using a tri­pod or flash in the pre­ced­ing years. With every new release, image sen­sor man­u­fac­tur­ers con­tin­ue to push their tech­no­log­i­cal advance­ments towards high­er ampli­fi­ca­tion while deliv­er­ing less elec­tron­ic noise than the out­go­ing mod­el. Despite this, it’s help­ful to under­stand what elec­tron­ic noise is, how it man­i­fests in your pic­tures, and what you can do to alle­vi­ate it, both dur­ing and after expo­sure.

Signal, noise, and the signal-to-noise ratio

Sig­nal: In pho­tog­ra­phy, light is the sig­nal record­ed by the image sen­sor to form a pho­to­graph. Specif­i­cal­ly, light is the por­tion of the elec­tro­mag­net­ic spec­trum that’s vis­i­ble to human eyes. Increas­ing the inten­si­ty of light by manip­u­lat­ing expo­sure pro­duces a stronger sig­nal that car­ries more infor­ma­tion.

Noise: Every ana­logue sys­tem designed to receive or trans­mit sig­nals will exhib­it some degree of noise. Noise is unde­sir­able because it inter­feres with and obscures por­tions of the sig­nal. It man­i­fests in sev­er­al vari­eties, all of which are unde­sir­able. There are gen­er­al­ly three broad caus­es of noise in dig­i­tal pho­tog­ra­phy. Shot noise is ran­dom noise caused by the uneven emis­sion of pho­tons from their source dur­ing any suf­fi­cient­ly short peri­od. Dark noise, like shot noise, is pro­duced by the inher­ent vari­a­tions in ther­mal­ly gen­er­at­ed elec­trons with­in the image sen­sor dur­ing a suf­fi­cient­ly short peri­od. Last­ly, there’s read noise, which is cre­at­ed by intrin­sic imper­fec­tions in the design of all elec­tron­ic com­po­nents that car­ry, inter­act with, and ampli­fy the sig­nal before its dig­i­ti­za­tion.

Read Noise may be reduced with bet­ter engi­neer­ing and greater pre­ci­sion in man­u­fac­tur­ing the elec­tron­ic com­po­nents. In sci­en­tif­ic appli­ca­tions, such as astron­o­my, dark noise is reduced by active­ly cool­ing image sen­sors to cryo­genic tem­per­a­tures. Since this isn’t fea­si­ble in prac­ti­cal pho­tog­ra­phy, man­u­fac­tur­ers work to reduce ther­mal gen­er­a­tion by design­ing their image sen­sors to be more ener­gy effi­cient, which reduces heat gen­er­a­tion, and by imple­ment­ing pas­sive cool­ing, such as inter­nal heat dis­si­pa­tors. Unfor­tu­nate­ly, shot noise is caused by an inher­ent prop­er­ty of light and can­not be reduced by cam­era design. The pur­pose of this semi-tech­ni­cal descrip­tion is to demon­strate that despite the seem­ing­ly unstop­pable trend of low-light high-ISO per­for­mance pur­sued by mod­ern cam­era man­u­fac­tur­ers, there are very real and very hard bar­ri­ers to this steady progress.

Sig­nal-to-Noise Ratio: The sig­nal-to-noise ratio, or SNR, is the amount of sig­nal divid­ed by the amount of noise. At lev­els above 1:1, there’s more sig­nal than noise in the sys­tem. A high­er SNR pro­duces results with less vis­i­ble noise. The only way to increase the SNR is by increas­ing expo­sure.

Types of image noise

Image sen­sors exhib­it three dif­fer­ent yet com­mon types of image noise: ran­dom, fixed-pat­tern, and band­ing.

Ran­dom Noise: This type of noise pro­duces sta­tis­ti­cal­ly ran­dom vari­a­tions in the bright­ness and colour of an image both above and below the signal’s lev­el. Although the pat­tern of ran­dom noise varies between pho­tos shot with iden­ti­cal expo­sure set­tings, the mag­ni­tude of the noise will remain the same. Ran­dom noise is present in all dig­i­tal images and is espe­cial­ly preva­lent in high ISO and short dura­tion expo­sures. It’s pri­mar­i­ly caused by shot noise and read noise.

Fixed-Pat­tern Noise: Fixed-pat­tern noise is often asso­ci­at­ed with the phe­nom­e­non of “hot pix­els,” which are pix­els that exhib­it greater inten­si­ty than both the sig­nal and the sur­round­ing ran­dom noise. While excep­tions occur, most fixed-pat­tern noise man­i­fests dur­ing long expo­sures and hot tem­per­a­tures. Hot pix­els and oth­er less pro­nounced types of fixed-pat­tern noise are gen­er­al­ly found in the same posi­tion across mul­ti­ple images when the shoot­ing con­di­tions are sim­i­lar.

Band­ing Noise: Band­ing is a type of semi-fixed-pat­tern noise that man­i­fests as faint ver­ti­cal or hor­i­zon­tal stripes. This type of noise is high­ly depen­dent on the spe­cif­ic image sen­sor employed in each cam­era. Some image sensors—and thus all cam­era mod­els fit­ted with them—are high­ly prone to band­ing noise gen­er­at­ed dur­ing the data trans­fer process. Band­ing is most evi­dent in the shad­ows of images shot with high ISO val­ues, or in under­ex­posed low ISO images that have been exces­sive­ly pushed.

Noise distribution across an image

Shad­ows, Mid-Tones, and High­lights: When exam­in­ing a pho­to that con­tains a vari­ety of tones, you’ll notice that the appar­ent dis­tri­b­u­tion and mag­ni­tude of noise isn’t entire­ly uni­form. The brighter parts of an image will exhib­it less vis­i­ble noise than the dark­er parts. This uneven dis­tri­b­u­tion of noise is caused by local­ized dif­fer­ences in the sig­nal-to-noise ratio. Since the brighter regions of an image receive more light, they record a high­er sig­nal and thus have a high­er SNR than the dark­er areas, which receive less light and record a com­par­a­tive­ly weak­er sig­nal.

Characteristics of random image noise

The noise from any colour-capa­ble dig­i­tal cam­era can be sep­a­rat­ed into two dis­tinct ele­ments: lumi­nance noise and chromi­nance noise.

Lumi­nance Noise: This presents as fluc­tu­a­tions in the tonal inten­si­ty of the noise pat­tern and is likened to film grain by many pho­tog­ra­phers.

Chromi­nance Noise: This presents as fluc­tu­a­tions in the hues of the noise pat­tern. Chromi­nance noise is unique to dig­i­tal pho­tog­ra­phy as there’s noth­ing anal­o­gous in film, and is com­mon­ly con­sid­ered unde­sir­able. The chief cause of chromi­nance noise is the dif­fer­ence in light trans­mis­sion between the three colours of the colour fil­ter array found on image sen­sors.

Noise reduction

In-Cam­era: Every cam­era offers some form of built-in noise reduc­tion algo­rithms. Most bridge, enthu­si­ast, and pro­fes­sion­al cam­eras will allow users to dis­able noise reduc­tion, or to spec­i­fy its strength. The algo­rithms are often broad in scope and apply even­ly to the entire image area. Exper­i­ment with vary­ing degrees of noise reduc­tion at dif­fer­ent ISO set­tings to see how it affects image qual­i­ty, specif­i­cal­ly the bal­ance between reduc­ing image noise and pre­serv­ing fine details and tex­tures. Last­ly, in-cam­era noise reduc­tion only affects the JPEG/TIFF out­put files; images record­ed in the RAW for­mat aren’t changed beyond the embed­ded pre­view.

Reduc­ing Lumi­nance and Chromi­nance Noise: When edit­ing images on a per­son­al com­put­er, some soft­ware appli­ca­tions allow users to sep­a­rate lumi­nance and chromi­nance noise reduc­tion into dis­crete process­es. Lumi­nance noise is gen­er­al­ly more chal­leng­ing to reduce, and the process tends to smooth out the detail in high-fre­quen­cy pat­terns, such as those of fab­rics, fur, and hair. Small to medi­um amounts of chromi­nance noise are gen­er­al­ly straight­for­ward to rem­e­dy on a per­son­al com­put­er. Although colour noise reduc­tion does not have a sig­nif­i­cant adverse impact on tex­tures, it will decrease the fine spa­tial res­o­lu­tion of colours.

Dark Frame Sub­trac­tion: Also known as long-expo­sure noise reduc­tion, this reduces fixed-pat­tern noise caused by long expo­sures. Imme­di­ate­ly fol­low­ing a long expo­sure, the cam­era takes a sec­ond and equal­ly long “expo­sure” with the shut­ter closed. The process records an image of the pre­vail­ing dark and read noise pat­terns and auto­mat­i­cal­ly sub­tracts it from the pre­vi­ous pho­to­graph. Dark frame sub­trac­tion must be imple­ment­ed man­u­al­ly when shoot­ing pho­tos in RAW for­mat.

Sharp­en­ing and Noise: In gen­er­al, sharp­en­ing an image will also sharp­en the image noise, there­by increas­ing its appar­ent mag­ni­tude. This is espe­cial­ly prob­lem­at­ic when in-cam­era noise reduc­tion is dis­abled and sharp­en­ing remains active.

AI Denoise Meth­ods: Mod­ern advance­ments in AI and machine learn­ing have led to the devel­op­ment of sophis­ti­cat­ed noise reduc­tion algo­rithms. For instance, Adobe’s “Denoise” func­tion in Light­room uses AI to ana­lyze and reduce noise effec­tive­ly while pre­serv­ing details. These AI-dri­ven tools can dif­fer­en­ti­ate between noise and actu­al image data, apply­ing tar­get­ed noise reduc­tion that min­i­mizes the loss of detail, a com­mon issue with tra­di­tion­al noise reduc­tion meth­ods.

Oth­er notable machine-learn­ing noise reduc­tion tools include Topaz DeNoise AI and DxO Pho­to­Lab’s Deep­PRIME. Topaz DeNoise AI employs deep learn­ing to reduce noise and enhance detail, mak­ing it effec­tive for both high ISO and low-light images. DxO Pho­to­Lab’s Deep­PRIME uses a com­bi­na­tion of denois­ing and demo­saic­ing pow­ered by AI, pro­vid­ing excel­lent results in noise reduc­tion and image qual­i­ty.

What’s too much noise?

Ulti­mate­ly, the impact of noise on your images is a mat­ter of per­son­al pref­er­ence. Some pho­tog­ra­phers find a bit of noise adds a cer­tain char­ac­ter to their pho­tos, while oth­ers strive to min­i­mize it as much as pos­si­ble. How much effort you put into reduc­ing noise—whether through care­ful expo­sure, post-pro­cess­ing tech­niques, or advanced AI tools—is entire­ly up to you and your cre­ative vision.