IMCO: Processing Color

How a camera works

Human Eye vs a Camera

• Shutter: how long the camera is opened to let in light

The imaging flow

The camera pipeline for creating nice photos is a secret recipe

The color Imaging Flow

Source

Other possible steps:

• ISO Gain
• Noise Reduction
• Etc.

Sensor

• Temporal Mutliplexing
  • Scan 3 times + use 1 sensor
  • 3 real values per pixel
  • Only for static scenes
  • Slow
• Scan 1 time + Use 3 sensors
  • 3 real values per pixel
  • Costly
  • Space
• Spatial Multiplexing
  • Scan 1 times + Use 1 sensor
  • Well-mastered technology
  • 1 real value per pixel (interpolation)
  • Loss of light

Camera spectral sensitivity

Source

Sensor (from 110 years ago!)

Using temporal multiplexing

Preprocessing

• Dark Current Compensation
  • Signal present even when the length is close, which adds noise

How to compensate ?

Capture a dark image for the given exposure time

$$C_i(x,y)=R_i(x,y)-D_i(x,y)$$

• Vignetting/Lens Shading/Flat Field correction
  • Image tends to darken on the corners/Non-uniform Illumination

How to compensate ?

You tell me !

Compute average RGB value of 15 White Patches

	Statistic	Red	Green	Blue
Before correction	Mean	157.3	159.1	157.5
After correction	Mean	249.5	\	\

If you are taking RAW images there is a chance your image software already does it

Color Constancy

Property of the Human Visual Sytem that allows adapting to different scene illuminations

White Balance to the rescue

The eye cares more about the intrinsic color of the object, not the color of the light leaving the object

• Easy for our eyes to judge what is white under different illuminants, but not so straightforward to camera

White Balance

The idea is to make white points the same between scenes

White point is the

$xyY$ (or

$XYZ$) value of an ideal "white reference"

Hence, the camera is able to do the Chromatic Adaptation

Chromatic Adaptation: How ?

• $\text{Source}$ color
  • $(X_s,Y_s,Z_s)$ with white reference
    $(X_{WS},Y_{WS},Z_{WS})$
• $Target$ color
  • $(X_T,Y_T,Z_T)$ with white reference
    $(X_{WT},Y_{WT},Z_{WT})$

Computing

$[M]$

1. Transform
   $XYZ$ to cone response domain
2. Scale using source/target white references
3. Transform back form cone domain
   $XYZ$

$$[M]=[MA{]}^{-1}\left[\begin{array}{ccc}\frac{{\rho }_T}{{\rho }_S}& 0& 0\\ 0& \frac{{\gamma }_T}{{\gamma }_S}& 0\\ 0& 0& \frac{{\beta }_t}{{\beta }_s}\end{array}\right][MA]$$

Where

$\rho ,\gamma ,\beta$ are the cone responses for the given source and target colors

Example

White balance: automatic

What if we don't help the camera ?

The camera doesn't know the illuminant

The camera will try to "guess" the white point in the image, and then balance the color automatically.

Unlike White Balance, the illuminant is very hard to estimate

How ?

• Gray world assumption
  • The average of all colors in the image is gray
  • Green channel is taken as "gray" reference
  • White-balanced image is:
    $k_{r}R,G,k_{b}B$
    • $k_r=\frac{G_{mean}}{R_{mean}}$
    • $k_b=\frac{G_{mean}}{B_{mean}}$
• White patch algorithm
  • Assumes that highlights = specular reflections of illuminant
  • Maximum
    $R,G,B$ values are good estimation of white point
  • White balanced image is:
    $k_{r}R,G,k_{b}B$
    • $k_r=\frac{G_{max}}{R_{max}}$
    • $k_b=\frac{G_{max}}{B_{max}}$
• How does GIMP does it ?
  1. Discards pixels colors at the extremes of
     $R,G,B$ histograms
     • Thresholds of
       $0.05\mathrm{\%}$ of total pixel count
  2. Do Histogram Stretching of remaining pixel colors, for each channel
  3. Plus some smart post-processing
• Smart Color Balance method
  1. Discards "black" and "white" pixels
  2. Stretch the histogram of the remaining pixels
  3. Plus some smart post-processing

Those are very basic algorithms.

Modern cameras use sophisticated white-balance, based on data-driven solutions

• One Way
  • Compute features from image
  • Find similar images in databases of images with good white balance
  • Use white balance from image

Demosaic

• Each pixel has a different color filter
• Bayer Pattern is the most used Color Filter Array (CFA)
• Why More Green ?
  • Frequency of the G color band is close to the peak of the human luminance frequency response
    $\to$ better sharpness

Linear interpolation

• Average of neighors
• Smoother kernels (bicubic) can also be used

Typical error

Taking a picture of striped pattern

$\to$ color artifact

To overcome these problems, most demosaicing methods convert the image to the YCbCr

Solution

• Each pixel has a different color filter
• Bayer Pattern is the moste used Color Filter Array (CFA)
• Interpolation methods are patented or proprietary
• And of course deep learning to the rescue

Color Transform

$RG{B}_c\to RG{B}_u$

Color Correction

• Cameras are meant to produce pleasing scenes rather than colorimetrically accurate scenes
  • Spectral Sensitivities of the camera are not identical to human color matching functions
• To obtain colorimetric accuracy, we need to transform the image from the sensor's color space to the colorimetrics
  $(XYZ)$ space
• By default Factory-computed Color Space Transforms (CST) are used
• When precision is needed, Color Charts are used to obtain a specific transform for the camera and scene

We can use ICC color profile

Color Correction Charts

ColorChecker Digital SG (SG=Semi-gloss)

• 140 patches (96 uniques colors)

Artist Paint Target

ColorBuild 300 Patch Target

Color Correction Methods

• $30+$ years of continous research on how to transform form RGB to XYZ spaces
  • A lot of ways to do it !
• Most of them solve the following equation:

$$X=MP$$

• $X$: Reference
  $XYZ$ values
  $[3\times m]$
• $P$: Camera
  $RGB$ values
  $[N\times m]$
• $M$: Correction matrix
  $[3\times N]$
• $m$: nb of data points
• $N$: nb of dimensions

	ColorChecker Classic	ColorChecker Digital SG
m	24	140

Linear

Linear Mapping from

$RGB$ to

$XYZ$

• $N=3$
• Not affected by exposure change

Polynomial Color Correction

Linear Mapping From

$RGB$ to

$XYZ$ + add polynomial components to reduce errors

• $N=9$, if
  $2^{nd}$ order polynomial
• Affected by exposure changes
• High polynomial degree tends to do overfitting

Root Polynomial Color Correction

Linear Mapping from

$RGB$ to

$XYZ$ + add root polynomial components

• $N=6$, if
  $2^{nd}$ Order Polynomial
• Not affected by exposure change

Measuring with RGB Cameras: Recap

• Cameras ar not light measurement devices
• From the moment an image is captured by the sensor, there are a lot of steps which could impact the final color rendition. This processing is proprietary

Future trends ?

White balance: Mixed illumination

The white balance fails if 2 illuminants are in the same image

Denoising

Best current solutions use Deep Learning

• Predict the noise instead of denoising the image

HDR Imaging

• HDR Imaging is quite mature
• Some challenges when acquiring moving scenes or shooting a video
• HDR from a single image using Deep Learning ?

Multispectral Imaging

• Used mostly for quality inspection and classification of materials
• In principle, it could be used for measuring Reflectance

Spectra from RGB

• When measuring color with an RGB camera, we measure the ligth response in 3 wavelengths

Color Science Toolbox