Noise Reduction in CMOS Image Sensors for High Quality Imaging The Autocorrelation Function Filter on Burst Image Sequences
Kazuhiro Hoshino (1), Frank Nielsen (2,3), Toshihiro Nishimura (4)
(1) Image Sensor Business Group, Sony Corporation, 4-14-1 Asahi-chou, Atsugi-shi, Kanagawa, Japan
(2) Sony Computer Science Laboratories, Inc. 3-14-13 Higashi Gotanda, Shinagawa-ku, Tokyo, Japan
(3) École Polytechnique, LIX, F-91128 Palaiseau Cedex, France
(4) Graduate School of Information, Production and Systems, Waseda University, 2-7 Hibikino, Wakamatsu, Kitakyushu, Fukuoka, Japan
1. Noise reduction in CMOS image sensors
for high quality imaging: The
autocorrelation function filter on burst
image sequences
Kazuhiro Hoshino1, Frank Nielsen2,3, Toshihiro Nishimura4
1 Image Sensor Business Group, Sony Corporation,
4-14-1 Asahi-chou, Atsugi-shi, Kanagawa, Japan
Kazuhiro.Hoshino@jp.sony.com,
2 Sony Computer Science Laboratories, Inc.
3-14-13 Higashi Gotanda, Shinagawa-ku, Tokyo, Japan
Frank.Nielsen@acm.org
3 Ecole Polytechnique, LIX F-91128 Palaiseau Cedex, France
4 Graduate School of Information, Production and Systems, Waseda University
2-7 Hibikino, Wakamatsu, Kitakyushu, Fukuoka, Japan
toshi-hiro@waseda.jp
2. Noise in image sensor
RST
Tx
Image pixel
P Offset noise(C)
N+
Reset noise(W)
N
SEL
Col. Bus 1/f noise(W )
Dark noise(C)
Dark shot noise( W )
n Photon shot noise( W )
Condensing Gm(C)
V Decoder
Amp noise( W)
Analog circuit
(Condenser, CDS, Decoder) Amp noise( W )
Offset noise(C)
Condensing Gm(C)
1/f noise( W)
TG Programmable
Gain Amp
CMOS image sensor W white noise, C colored noise.
3. Principle of an ACF
The data is collected at the same interval time.
Autocorrelation value is calculated according to the following equation.
N 1
1
R( ) x(t ) x(t )
N t 0
R is ACF value.
N is the number of data,
t is time.
x is pixel value,
and τ is shifted time.
4. 1D simulation of ACF
(A) cosine wave
Block diagram of 1-D ACF method
(B) white noise wave
Sampling
Make Calculation
+ In same
base wave ACF value
interval time
Make
noise wave
white noise wave
Original
wave
ACF
value
(A) cosine wave (B) white noise wave
5. Expansion ACF method to 2-D model
Time
V
Image
H H direction
N 1
1
R( ) x(t ) x(t )
N t 0
R is ACF value.
N is the number of data which were sampled in time axis,
t is time.
x is pixel value,
and τ is shifted time.
6. ACF value as a function of pixel intensity
(A)
Auto Correlation Value Pixel-A
(B)
Pixel-B
Flame Number
Bright pixel A (180 in 256 scale) and dark pixel B (8 in 256 scale)
7. The algorithm of noise judging and filtering process
by a time domain ACF method
Image data (BMP,RAW)
Pixel value extraction
Pixel value i<10 No
Calculation of ACF
ACF value r<0.8 No
Leveling filter processing
Pixel value decision
I< Total pixel number
No
END
8. Result of image processing
・ Reduction of random noise is possible per pixel.
・ Since filter processing is not performed in a bright pixel, resolution does not
deteriorate.
Original image Processing image
9. The algorithm and the example of processing of a
time domain ACF method
Image data (BMP,RAW)
Pixel value extraction
Pixel value i<10 No
Calculation of ACF
ACF value r<0.8 No
Leveling filter processing
Pixel value decision
I< Total pixel number
No
END
10. Image processing result as a function of threshold
value both pixel value and ACF value
Original Ith= 100
Rth=0.985
Ith= 100 Ith= 100
Rth=0.995 Rth=1.000