You are on page 1of 4

A Novel Smoothness-Based Interpolation Algorithm

for Division of Focal Plane Polarimeters

Jieyun Zhang1, Wenbin Ye1 , Ashfaq Ahmed2, Zhurui Qiu3 , Yuan Cao1 and Xiaojin Zhao1,4,∗
1 College of Electronic Science and Technology, Shenzhen University, Shenzhen, China
2 Department of Bioengineering, the Hong Kong University of Science and Technology, Hong Kong
3 Chenghan International School, Shenzhen, China
4 American University of Ras Al-Khaimah, Ras Al-Khaimah, U.A.E.
∗ Email: eexjzhao@szu.edu.cn

CMOS Image Sensor Column Selection


Abstract—In this paper, we present a novel smoothness-
based interpolation algorithm for the division of focal plane 0 0 0 0 1 1

CMOS Image Sensor Row Selection


polarimeters (DoFP). By calculating the divided blocks’ variance
that represents their local smoothness, the proposed algorithm

1 1 0 0 0 0
well-balances between the traditional bilinear and bicubic inter- micropolarizer for 0r linearly
0º Micro-polarizer
polarized component
polation algorithms. In addition, compared with the previously
reported gradient-based interpolation algorithm which only indi- micropolarizer for 90r linear
cates the image’s directional change along 0◦ , 90◦ , 45◦ and 135◦ , 90º Micro-polarizer
polarized component
the presented smoothness-based algorithm covers the variations
along all the possible directions, leading to more accurate se- micropolarizer for 45r linear
45º Micro-polarizer
polarized component
lection between the bilinear and bicubic algorithms. According
to our extensive simulation results, the proposed implementation micropolarizer for 135r linea
exhibits the lowest mean square error (MSE) for the test images 135º Micro-polarizer
polarized component
among all the previously reported algorithms, including bilinear,
bicubic and gradient-based interpolation algorithms.

I. I NTRODUCTION Fig. 1. Typical 2×2 micro-polarizer pattern for DoFP sensors.

Recent advancement in the nanofabrication field has en-


abled the promising concept of division of focal plane po-
larimeters (DoFP), which is also known as the polarization the micro-polarizers on the imaging array are oriented in
imaging based on the focal plane micro-polarimeters with four different directions and each records one of the four
predefined pattern [1]–[5]. Similar to the color filter array polarization signatures. This results in down-sampled sub-
(CFA) widely exploited in the color image sensor domain [6], images and the reduction of spatial resolution. In order to
traditional monochromatic image sensors can be endowed with improve both the spatial resolution and the accuracy of the
the capability of polarization imaging by integrating a layer polarization information, it is necessary to interpolate the
of pixelated micro-polarizers on top [1]–[5]. Consequently, images obtained from DoFP polarization sensors. Currently,
as shown in Fig. 1, in most literatures, the micro-polarizers several interpolation algorithms exist, including bilinear, bicu-
are arranged to a 2×2 periodic pattern, with their four local bic and gradient-based algorithms [9]–[13]. The bilinear algo-
major polarization axes oriented typically along 0◦ , 90◦ , 45◦ , rithm has the lowest computational complexity but generally
135◦ [1]–[3], [5]. In the CFA-based color image sensors, it yields the worst results especially when dealing with images
is well-known that: in order to have the different wavelength having complex spatial information [9]. The bicubic algorithm
channels (i.e. R, G and B) simultaneously extracted, the spatial may produce better results as indicated in [9], but suffers
resolution of each channel needs to be compromised. As from the high computational complexity. In order to alleviate
a result, different interpolation algorithms are necessary to this limitation of computation complexity, the gradient-based
compensate the spatial resolution loss, namely “demosaicing” interpolation algorithm is introduced in [10], [11]. In this
[6], [7]. Here for the DoFP-based polarization imaging, we algorithm, the gradient is calculated for 0◦ , 90◦ , 45◦ and
can capture different polarization channels in one single frame, 135◦ directions, respectively. If the gradient is greater than
however, the polarization image sensors suffer from the similar a threshold value, the high-order interpolation algorithm (i.e.
problem of spatial resolution loss and low accuracy of captured bicubic interpolation algorithm) is used. Otherwise the bilinear
polarization images, hence the same concept of “polarization interpolation algorithm is adopted. However, the gradient is an
demosaicing” applies as well [8]. indicator of the directional change. It is straightforward that
the directional change can occur in any direction. Thus the
Since the image is sampled through four pixelated micro- algorithm proposed in [10], [11], which only calculates the
polarizers scattered throughout the space, the pixels in the gradient in four directions, may not produce the most accurate
aforesaid 2×2 periodic pattern record different polarization results.
information of the instantaneous field of view. Therefore, if
we were to directly calculate the polarization related physi- In this paper, a novel algorithm that aims to improve
cal parameters according to those polarization signatures, an the gradient-based interpolation algorithm is proposed. Instead
error is introduced as there will be a deviation from the of using the gradient to represent the image variations, the
true polarization information of the target [8]. Additionally, smoothness is exploited. If the smoothness variation of the
f(x,y) f(x+1,y) f(iͲ1,jͲ1) f(i,jͲ1) f(i+1,jͲ1) f(i+2,jͲ1)

f(u,v) f(iͲ1,j) f(i,j) f(i+1,j) f(i+2,j)

f(i+u,j+v)

f(x,y+1) f(x+1,y+1)
f(iͲ1,j+1) f(i,j+1) f(i+1,j+1) f(i+2,j+1)

Fig. 2. The traditional bilinear interpolation algorithm.


f(iͲ1,j+2) f(i,j+2) f(i+1,j+2) f(i+2,j+2)

image is larger than a threshold value, the high-order interpo-


lation algorithm (i.e. bicubic) is used; otherwise, the bilinear Fig. 3. The traditional bicubic interpolation algorithm.
algorithm is adopted. The rest of this paper is structured as
follows: Section II covers a review of the bilinear and bicubic
interpolation algorithms; the proposed smoothness-based in-
terpolation algorithm is presented in Section III; Section IV The formula of bicubic interpolation is expressed as follows:
shows the results and compares the proposed interpolation f (i + u, j + v) = A · B · C (3)
algorithm to the previous ones; finally, the paper is concluded
in Section V. where A, B and C are matrices expressed as follows:
T
II. R EVIEW OF THE TRADITIONAL INTERPOLATION A = [ s(1 + u) s(u) s(1 − u) s(2 − u) ] (4)
ALGORITHMS FOR D O FP
2 3
f (i − 1, j − 2) f (i, j − 2) f (i + 1, j − 2) f (i + 2, j − 2)
A. Bilinear Interpolation Algorithm 6 f (i − 1, j − 1)) f (i, j − 1) f (i + 1, j − 1) f (i + 2, j − 1) 7
B=4 f (i − 1, j) f (i, j) f (i + 1, j) f (i + 2, j) 5
The principle of the bilinear interpolation algorithm ex- f (i − 1, j + 1) f (i, j + 1) f (i + 1, j + 1) f (i + 2, j + 1)
tends the one-line interpolation method to two-variable in
a two-dimensional grid. Specifically, the pixel value to be
T
interpolated is calculated by its four neighboring pixels with C = [ s(1 + v) s(v) s(1 − v) s(2 − v) ] (5)
different weights. From Fig. 2, to determine the pixel value
of the interpolated point f (u, v), the four neighboring pixel
points f (x, y), f (x + 1, y), f (x, y + 1), and f (x + 1, y + 1) III. P ROPOSED S MOOTHNESS -BASED I NTERPOLATION
are adopted as follows: A LGORITHM
f (u, v) = f (x, y) ∗ (x + 1 − u) ∗ (y + 1 − v) As above-mentioned, the gradient is an indicator of the
+f (x, y + 1) ∗ (x + 1 − u) ∗ (v − y) directional change, which can occur in any direction. The pre-
viously reported implementation only calculates the gradient
+f (x + 1, y + 1) ∗ (u − x) ∗ (v − y) in four directions [10], [11], which can be further improved
+f (x + 1, y) ∗ (u − x) ∗ (y + 1 − v) (1) to cover the directional change in other orientations. In this
section, in order to further elevate the interpolation accuracy of
It is known that the computational complexity is low for the the gradient-based algorithm [10], [11], we proposed to utilize
bilinear interpolation, however, the interpolation accuracy may the smoothness for selecting the more appropriate algorithm.
not be excellent for different image conditions. The fundamental idea behind the proposed method is: if the
smooth region of the image is encountered, the bilinear method
B. Bicubic Interpolation Algorithm is used; otherwise, the bicubic method is adopted. So the
key factor of the proposed method is how to compute the
Bicubic interpolation, also known as the cubic convolution smoothness of the image. In this paper, the pixels’ variance of
interpolation, considers 16 neighboring pixels (4×4) (Fig. 3), an image’s selected region is used to represent the smoothness.
which is different from the bilinear interpolation only taking The proposed method is described using a flow chart in Fig.
4 pixels (2×2) into account. For the image pattern with 4, which mainly involves the following steps:
large image variations, this interpolation algorithm generally
yields superior results, however, it is at the expense of high 1) Divide the given polarization image into small blocks,
computational complexity. This can be reflected by the base here the block size is set to be 2×2.
function of the bicubic interpolation: 2) Evaluate the smoothness of image blocks by com-
⎧ ⎫ puting their variances, respectively. The 2×2 block’s

⎪ 1 − 2|W |2 + |W |3 |W | < 1 ⎪ ⎪ variance σ is derived as follows:

⎨ ⎪

s(W ) = 2 3
4 − 8|W | + 5|W | − |W | 1  |W | < 2 1
⎪ ⎪
(2) J = [(f (x, y) − E)2 + (f (x + 1, y) − E)2

⎪ ⎪
⎪ 3
⎩ ⎭ +(f (x, y + 1) − E)2 + (f (x + 1, y + 1) − E)2 ] (6)
0 |W |  2
method can be executed with much lower computational
complexity. Regarding the gradient-based method [11], it is
difficult to have a fair comparison in term of computation
complexity. The reason is that the gradient-based method and
the smoothness-based method are two different criterias to sep-
arate pixels according to the interpolation algorithm adopted
(i.e. bilinear or bicubic), which can yield very different pixel
separation results with the same image, and is quite image-
dependent. Therefore, we choose to compare the accuracy of
the interpolation algorithm in the following section.

IV. R ESULTS AND D ISCUSSIONS


In order to effectively evaluate the performance of different
interpolation algorithms, high resolution polarization images
were obtained by rotating a linear polarization filter in front of
an 8-bit grey-scale CCD camera. These original full-resolution
images, which are not available even with the real DoFP
sensors, are adopted here as the ground truth for evaluating
different interpolation algorithms. Then they are down-sampled
and interlaced to generate the test images, which can well-
mimic the polarization images generated by the real DoFP
sensors. Successively, the four interpolation algorithms (i.e. bi-
linear, bicubic, gradient-based, smoothness-based) are applied
to these generated test images to compute the full-resolution
images. After that, these interpolated images are compared
with the ground-truth images in term of the mean square error
(MSE), which is commonly-used to evaluate the performance
of different interpolation algorithms [9]–[11]. Here the MSE
is defined by:
1
M SE = [Oc (i, j) − Ic (i, j)]2 (8)
MN
1iM 1jN

where M ×N is the test image’s array size, Oc (i, j) is the pixel


value of the ground truth polarization images, and Ic (i, j) is
the interpolated polarization images.
Generally speaking, lower MSE corresponds to higher
accuracy of the applied interpolation algorithm. As a result,
as shown in Table I, the proposed smoothness-based algorithm
Fig. 4. Flow chart of the proposed smoothness-based interpolation algorithm. is compared with the previously reported bilinear, bicubic and
the gradient-based algorithms, in term of MSE. It is observed
that among the four algorithms, the bilinear interpolation
√ method generates the results with the largest MSE for all four
σ= J (7) polarization sub-images; while the proposed smoothness-based
where E is mean value of the block, f (x, y), f (x + method generates the results with the minimum MSE for all
1, y), f (x, y + 1), f (x + 1, y + 1) are four provided four polarization sub-images. Meanwhile, it is noted that even
pixel values. though the bicubic interpolation method has higher interpola-
3) Compute the threshold value σmedian . Here the tion order (i.e. computation compexity), it does not necessarily
threshold value is the median value of all the blocks’ leads to lower MSE or higher interpolation accuracy. In other
variances. words, the interpolation accuracy is not closely related to the
4) Compare each block’s variance σ to the threshold interpolation order. For example, we have a test image in an ex-
value σmedian . If σ < σmedian , it indicates the block tremely smooth case where all the pixels are equal to a constant
can be considered as a smooth region, and the bilinear value. As a result, the best interpolation method should be the
interpolation method is applied to this block; while low order interpolation method (e.g. zero-hold interpolation),
if σ > σmedian , the block is considered as a non- not the high-order bicubic interpolation. Therefore, if an image
smooth region, and the bicubic interpolation method is smooth enough, the bilinear method may lead to lower
is applied. MSE than the bicubic method. The proposed smoothness-
based interpolation method exhibits the highest accuracy since
It is known that 1 multiplication and 3 additions are re- it takes the smoothness of an image’s different blocks into
quired for calculating E, and 6 multiplications plus 7 additions account. If the block of an image is relatively smooth, which
are required for calculating σ. By adding 1 multiplication and mainly corresponds to the low frequency components, the
3 additions for the bilinear interpolation of a 2 × 2 window, low order interpolation method (i.e. bilinear) is preferred in
we have altogether 8 multiplications and 13 additions. Com- our case; otherwise, the bicubic method is adopted. Fig. 5
pared with the bicubic method requiring 70 multiplications presents one set of the test images interpolated by four different
and 45 additions, the proposed smoothness-based interpolation algorithms.
TABLE I. MSE COMPARISON WITH THE PREVIOUS INTERPOLATION ALGORITHMS .

Bilinear Bicubic Gradient-based Smoothness-based


MSE of 0◦ 8.7423 5.7530 6.7545 5.6796
MSE of 90◦ 9.8282 6.5321 7.6749 6.4585
MSE of 45◦ 9.1837 6.1072 7.1235 6.0389
MSE of 135◦ 9.3820 6.1806 7.3626 6.1039

Bilinear Bicubic GradientͲBased SmoothnessͲBased

0e

(a) (b) (c) (d)

45e

(e) (f) (g) (h)

90e

(i) (j) (k) (l)

135e

(m) (n) (o) (p)

Fig. 5. One set of the test images interpolated by the four different algorithms.

V. C ONCLUSIONS [3] M. Kulkarni and V. Gruev, “Integrated spectral-polarization imaging


sensor with aluminum nanowire polarization filters,” Opt. Express, vol.
In this paper, a smoothness-based algorithm is proposed to 20, no. 21, pp. 22997–23012, 2012.
further improve the interpolation accuracy of the DoFP sen- [4] K. Sasagawa, S. Shishido, K. Ando, H. Matsuoka, T. Noda, T. Tokuda,
sors. Compared with all the previously demonstrated interpo- K. Kakiuchi and J. Ohta, “Image sensor pixel with on-chip high extinc-
lation algorithms for polarization images, the proposed imple- tion ratio polarizer based on 65-nm standard CMOS technology,” Opt.
Express, vol. 21, no. 9, pp. 11132–11140, 2013.
mentation exhibits the highest accuracy with lowest MSE for
[5] T. York, R. Marinov and V. Gruev, “260 frames-per-second 648x488
all different polarization channels. Moreover, the algorithm’s resolution division-of-focal-plane polarimeter with structural dynamics
overall computation complexity is well-balanced between the and tracking applications,” Opt. Express, vol. 24, no. 8, pp. 8243–8252,
traditional mainstream bilinear and bicubic algorithms. This 2016.
makes it quite suitable for the on-chip hardware implemen- [6] B. E. Bayer, “Color imaging array,” US Patent 3,971,065, 1976.
tation, which enables the promising concept of polarization [7] O. Losson, L. Macaire, Y. Yang, “Comparison of Color Demosaicing
image sensing and processing together on a single chip. Methods,” Advances in Imaging and Electron Physics, vol. 162, pp. 173–
265, 2010.
[8] B. M. Ratliff, C. F. LaCasse and J. Scott Tyo, “Interpolation strategies for
ACKNOWLEDGMENT reducing IFOV artifacts in microgrid polarimeter imagery,” Opt. Express,
vol. 17, no. 11, pp. 9112–9125, 2009.
This work was supported by the National Natural Sci- [9] S. Gao and V. Gruev, “Bilinear and bicubic interpolation methods for
ence Foundation of China (Grant No. 61601168, 61504087), division of focal plane polarimeters,” Opt. Express, vol. 19, no. 27, pp.
ICT Fund of UAE, the Fundamental Research Founda- 26161–26173, 2011.
tion of Shenzhen (Grant No. JCYJ20150324141711677, [10] S. Gao and V. Gruev, “Gradient Based Interpolation for Division of
Focal Plane Polarization Imaging Sensors,” in Proc. IEEE International
JCYJ20160520170741660, JCYJ20160308094919279) and the Symposium on Circuits and Systems (ISCAS), pp. 1855–1858, 2012.
Natural Science Foundation of SZU (Grant No. 2016054). [11] S. Gao and V. Gruev, “Gradient-based interpolation method for division-
of-focal-plane polarimeters,” Opt. Express, vol. 21, no. 1, pp. 1137–1151,
2013.
R EFERENCES [12] J. Zhang, H. Luo, B. Hui and Z. Chang, “Image interpolation for
[1] V. Gruev and R. Perkins, “A 1 MPixel CCD Image Sensor with division of focal plane polarimeters with intensity correlation,” Opt.
Aluminum Nanowire Polarization Filter,” in Proc. IEEE International Express, vol. 24, no. 18, pp. 20799–20807, 2016.
Symposium on Circuits and Systems (ISCAS), pp. 629–632, 2010. [13] E. Gilboa, J. P. Cunningham, A. Nehorai and V. Gruev, “Image
[2] V. Gruev, R. Perkins and T. York, “CCD polarization imaging sensor interpolation and denoising for division of focal plane sensors using
with aluminum nanowire optical filters,” Opt. Express, vol. 18, no. 18, Gaussian processes,” Opt. Express, vol. 22, no. 12, pp. 15277–15291,
pp. 19087–19094, 2010. 2014.

You might also like