Professional Documents
Culture Documents
Implementation of An Image Signal Processor For Reconfigurable Processors
Implementation of An Image Signal Processor For Reconfigurable Processors
Implementation of An Image Signal Processor For Reconfigurable Processors
net/publication/261635313
Conference Paper in Digest of Technical Papers - IEEE International Conference on Consumer Electronics · January 2014
DOI: 10.1109/ICCE.2014.6775944
CITATIONS READS
3 674
4 authors:
Some of the authors of this publication are also working on these related projects:
Research on Multimedia SoC and Platform for Next Generation TV View project
All content following this page was uploaded by Junguk Cho on 26 September 2018.
Abstract—An image signal processor for a CMOS image demosaicing algorithm. The AHD consist of three steps;
sensor consists of many complicated functions. In this paper, the directed interpolation, homogeneity-directed map creation,
full chain of camera ISP functions for smart devices is presented. iteration. The directed interpolation is made of fixed point
The every function in the chain is fully converted to fixed point
arithmetic with 2bit extension. The coefficient of
arithmetic and no special function is used for easy porting to
reconfigurable processors. To verify the performance of the homogeneity-directed map creation is converted to fixed
proposed ISP chain, series of test are performed, and all point. Also, CIELab that is used as the color space in the
measured values satisfied the quality requirements. original homogeneity-directed map is changed to YCbCr color
space. The iteration function is removed for reducing
I. INTRODUCTION operational loads. After demosaicing, the color correction
Recently, the technology progress of smart devices is block finds color features and repairs color artifacts.
surprisingly fast. Among many performance improvements Linear stretch method is used in the auto contrast. The
and additional special purpose functions, one of the most linear scale factors in the auto contrast function are calculated
commercially successful functions is the digital camera. in YPbPr color space. A color control function is also used in
Nowadays almost every smart device has the digital camera YPbPr color space to control color saturation and color offset.
feature. Since the environment where the camera with the Gamma Correction with a reduced LUT(LookUp Table) is
smart devices is used is sometimes not quite suitable to get used and the gamma values between LUT entries are
good quality pictures. Therefore, Bayer pattern image taken calculated by piecewise linear interpolation method. [3]
from the CMOS image sensor used in the digital camera Modified BF(Bilateral Filter) is used as a noise reduction
should be converted to the RGB image using many complex algorithm and based on BF[2]. The original BF has two
enhancement functions in the ISP (Image Signal Processor). Gaussian filters. One is for distance weight between pixel
locations. The other is for difference weight between pixel
II. ALGORITHM intensities. In order to simplify the two Gaussian filters, the
Gaussian functions are replaced by fixed point binary
In this paper, the full chain of camera ISP functions for
threshold functions. The threshold values are determined by
smart devices is presented. The every function of the chain is
pre-calculating the Gaussian filter coefficient for pixel
fully converted to fixed point arithmetic and no special
locations and pixel intensities.
function is used for easy porting to reconfigurable processors.
For detail enhancement DoG(Difference of Gaussian) based
Fig.1 is flow chart of ISP full chain for vector processors.
LTI/CTI[4] is used. The Gaussian mask sizes in the LTI/CTI
As in Fig. 1, after performing white balance in the bayer
are 3x3 and 5x5, which are with pre-calculated coefficients.
pattern, Modified AHD(Adaptive Homogeneity-directed
Demosaicing), which is based on AHD[1], is used as
Requirements (20lux) 0.33 ~ 0.8 20% 0.65 -1.2 ~ 0.5 0.4 ~ 0.9 30 dB
Requirements (200lux) 0.35 ~ 0.7 20% 0.15 pixel 0.65 -0.5 ~ 0.5 0.4 ~ 0.7 30 dB
Results (20lux) 0.545 0% 0.753 -0.81 0.4 39.23dB
Results (200lux) 0.737 0% 0.062 pixel 0.758 -0.49 0.56 47.03dB
Metric DR Light Falloff Color Uniformity SMIA Delta C00
2900K 3500K 5500K
Requirements (20lux) 33 dB Mean 15 Mean 10
Max 20 Max 15
Requirements (200lux) 33 dB 70% RI Max ෙ C 10 |SMIA| < Mean 10 Mean 10
130% 6% Max 15 Max 15
Results (20lux) 44dB Mean 8.1 Mean 5.6
Max 13.6 Max 12.6
Results (200lux) 44dB 72.43% 2.23 -2.55 Mean 4.8 Mean 4.7
Max 14.0 Max 14.9