Modelling, calibration and rendition of colour logarithmic CMOS image sensors Dileepan Joseph and...
-
date post
19-Dec-2015 -
Category
Documents
-
view
214 -
download
0
Transcript of Modelling, calibration and rendition of colour logarithmic CMOS image sensors Dileepan Joseph and...
Modelling, calibration and rendition of colour logarithmic CMOS image sensorsDileepan Joseph and Steve CollinsDepartment of Engineering ScienceUniversity of Oxford
IMTC 2002 2May 21-23
Outline
Logarithmic CMOS image sensors Modelling sensor response Image sensor calibration
Fixed pattern noise Sensation of colour
Rendition of images CIE Lab (perceptual error) IEC sRGB (standard display)
Summary and future work
IMTC 2002 3May 21-23
Logarithmic CMOS image sensors
CMOS displacing CCD because of integration of signal processing and economies of scale
Logarithmic sensors offer high dynamic range and high frame rate
Linear sensors offer low fixed pattern noise and good colour rendition
Example images taken from IMS Chips website
Linear CCDsensor
Logarithmicsensor
Logarithmicsensor
IMTC 2002 4May 21-23
Modelling sensor response
Since Ik = ∫ fk(λ) s(λ) dλ For photocurrent Ik, spectral response
fk(λ) and light stimulus s(λ) at a pixel, where k = R, G or B
And fk(λ) = gL(λ) gk(λ) gP(λ) For spectral responses of lens gL(λ),
colour filter gk(λ) and photodiode gP(λ) Approximating a linear combination of
three CIE XYZ basis functions Then Ik = dk • x
For mask coefficients dk and tricolour vector x, i.e. s(λ) in CIE XYZ space
Ideally, y = a + b ln (c + Ik) + ε For digital response y of pixel with
offset a, gain b, bias c and error ε Pixel-to-pixel variation of a, b or c
causes fixed pattern noise (FPN)
IMTC 2002 5May 21-23
FPN calibration
Three types of FPN of interest: Offset variation Offset and gain variation Offset, gain and bias variation
Partition pixels by colour filter to permit FPN calibration of three monochromatic sensors
Take images of uniform stimuli under different illuminances
Calibrate each pixel’s response to average response of all pixels by least squares estimation of varying model parameters
Fuga 15RGB sensor exhibits offset, gain and bias variation
IMTC 2002 6May 21-23
Colour calibration
Take and segment images of a standard chart, having patches of known CIE XYZ colour, under different illuminances
Calibrate pixel responses to colour by estimating non-varying model parameters (e.g. mask dk), using estimates of varying parameters
Ideal model fails for Fuga 15RGB because absolute relationship between y and Ik invalid (strong inversion component?)
Empirical model y = a + b ln (c + (α + dk • x)β) worked well, with no change to relative responses of pixels or FPN calibration
IMTC 2002 7May 21-23
Image rendition (CIE Lab)
Images of a Macbeth Colour Chart, taken by the Fuga 15RGB, were rendered into CIE Lab space with the calibrated empirical model
The perceptual error increases in dim lighting as the bias term c dominates the photocurrent Ik
Excluding the dimmest image (i.e. 5 lux), the error equals 12 over a 60 dB dynamic range for offset, gain and bias variation
Images in Digital Photographer show that conventional (linear) digital cameras have an error of 15 over a 30 dB dynamic range
IMTC 2002 8May 21-23
Image rendition (IEC sRGB)
A Fuga 15RGB image of the Macbeth Chart, taken in 11 lux of illuminance, was rendered into IEC sRGB space with the calibrated empirical model
Results for offset variation (top-left), offset and gain variation (top-right), offset, gain and bias variation (bottom-left) and true colours (bottom-right) are shown
Two types of residual deviation for the rendered patches are visible: Fixed pattern noise Colour desaturation
IMTC 2002 9May 21-23
Summary and future work
Logarithmic image sensors offer high dynamic range and frame rate Combine theories of colour linear sensors and monochromatic
logarithmic sensors to model colour logarithmic sensors Calibrate FPN, using images of uniform stimuli, by relative
estimation of model parameters that vary from pixel to pixel Calibrate colour, using images of a colour chart, by absolute
estimation of model parameters that do not vary Fuga 15RGB results expose limitations of ideal model in absolute
estimation but reveal empirical model that works well Macbeth Chart results show colour rendition with calibrated Fuga
15RGB competes with conventional digital cameras Seek to minimise bias variation, so simple FPN models suffice, and
bias magnitude, to improve colour rendition in dim lighting
IMTC 2002 10May 21-23
Acknowledgements
The authors gratefully acknowledge the support of the Natural Sciences and Engineering Research Council (Canada) and the Engineering and Physical Sciences Research Council (UK)