IEICE TRANS. INF. & SYST., VOL.E90–D, NO.

2 FEBRUARY 2007
606

LETTER
A Low-Complexity Interpolation Method for Deinterlacing

Pei-Yin CHEN†a) , Member and Yao-Hsien LAI† , Nonmember

SUMMARY A direction-oriented spatial interpolation technique for proposed in [6].

image de-interlacing is presented in this letter. The experimental results In this letter, we present an efficient and low-
demonstrate that our method achieves excellent performance in terms of
complexity interpolation method. The proposed method in-
both objective and subjective image quality. The proposed algorithm also
has a very computationally simple structure, and proves to be a good can- troduces additional measurements for estimating the spatial
didate for low-cost hardware interpolator. correlations efficiently, with the primary objective of alle-
key words: interpolation, low-cost viating decision errors for subsequent interpolation. Com-
pared with the previous methods, our method can detect
1. Introduction edges efficiently and perform better in terms of both objec-
tive and subjective image quality. Furthermore, the method
Various digital video systems need different spatial and tem- has very low computational complexity, so it can be easily
poral resolutions, so different format standards are used to implemented with hardware and applied to many real-time
store, transmit and display digital video signals. Traditional applications.
NTSC-TV uses the interlaced scanning format to display The rest of this letter is organized as follows. In Sect. 2,
video sequences. It suffers from uncomfortable visual ar- the proposed method is described in detail. The simulation
tifacts such as edge flicker, interline flicker and line crawl- results are provided in Sect. 3. Finally, conclusions and re-
ing. Recently, many progressive scanned devices such as marks are given in Sect. 4.
HDTV, PDPs and LCD-TVs become more and more popu-
lar. Hence, an efficient image de-interlacing (or conversion) 2. The Proposed Method
technique is necessary to convert the interlaced images into
progressive images for suitable displaying. The motivation of the proposed interpolation algorithm is
Interpolation is used widely in image de-interlacing to first to estimate efficiently the spatially directional correla-
obtain higher image resolution [1]. In the past few years, tions, and then to interpolate the missing pixels accordingly.
many image interpolation methods have been proposed, Let x(i, j) denote the signal to be interpolated where i means
ranging from spatial techniques [2]–[6] to more complex the number of horizontal line and j means the number of
temporal techniques [7]–[10]. Temporal techniques usually vertical line. The signals x(i − 1, j) and x(i + 1, j) are sup-
perform better than spatial techniques with regard to image posed to be available and can be used as the inputs for in-
quality, but they require more expensive computations and terpolation. For the purpose of checking edges existed in
larger memory space. In contrast, the spatial technique is the diagonal, vertical and horizontal directions respectively,
very suitable for real-time applications due to its simplicity four directional differences, denoted as Dd1 , Dd2 , Dv and Dh ,
and low computational requirements. In this letter, we con- are defined and calculated as follows:

sider only low-complexity spatial interpolation techniques. 

 Dd1 = |x(i − 1, j − 1) − x(i + 1, j)|



Many spatial interpolation techniques have been pro- 
 + |x(i − 1, j) − x(i + 1, j + 1)|



posed for image enlargement. In [3], an edge-based line 
 D d2 = |x(i − 1, j) − x(i + 1, j − 1)|

average (ELA) interpolation method was developed. Chen 
 + |x(i − 1, j + 1) − x(i + 1, j)| (1)



et al. [4] presented an efficient algorithm using two useful 
 D = |x(i − 1, j) − x(i + 1, j)| × 2



v
measurements defined on the sliding window to increase the 
 Dh = |x(i − 1, j − 1) − x(i − 1, j)|


accuracy of the ELA algorithm. In [5], Kim et al. employed + |x(i + 1, j − 1) − x(i + 1, j)|
an adaptive pseudomedian filter to implement an image in-
terpolator. Based on the median filter [1] and the modified Figure 1 shows the four pixel differences of different direc-
ELA method [4], a median-based interpolation method is tions, respectively. The smaller one among these four values
has a stronger correlation, and probably has an edge in its
Manuscript received May 22, 2006. direction. Hence, the interpolated pixel can be estimated as
Manuscript revised August 2, 2006. follows:
†
The authors are with the Department of Computer Science
and Information Engineering, National Cheng Kung University,
Tainan 701, Taiwan, R.O.C.
a) E-mail: pychen@csie.ncku.edu.tw
DOI: 10.1093/ietisy/e90–d.2.606

Copyright

c 2007 The Institute of Electronics, Information and Communication Engineers
LETTER
607




 x(i, j − 1), if Dh = 0


 − 1, j − 1) + x(i + 1, j) +



(x(i 3. Simulation Results


 x(i − 1, j) + x(i + 1, j + 1))/4,





 if min(D d1 , Dd2 , Dv ) = Dd1 For an M × N image, we remove every odd vertical line

x(i, j) = 
 (x(i − 1, j) + x(i + 1, j − 1) + (2) and thus obtain the M/2 × N image. Then, we reconstruct




 x(i − 1, j + 1) + x(i + 1, j))/4, the M × N image by interpolating the odd lines for differ-




 if min(D , Dd2 , Dv ) = Dd2 ent methods. Totally, ten 512 × 512 standard test images



d1

 (x(i − 1, j) + x(i + 1, j))/2,

 if min(D , D , D ) = D with 8-bit brightness resolution are considered. The results
d1 d2 v v of peak signal-to-noise ratio (PSNR) are illustrated in Ta-
ble 1. Except our method, five previous interpolation meth-
Obviously, the proposed method requires only simple in-
ods, FOI [2], ELA [3], ELC [4], APM [5] and MBI [6], are
teger operations, such as addition, subtraction and abso-
used for comparisons. It can be observed from the results
lute value, so it can be easily implemented with hard-
that our algorithm achieves better quantitative quality than
ware. The simulation results shown in the following section
the previous interpolation methods. Certainly, the exact de-
will demonstrate that our method achieves excellent perfor-
gree of improvement is dependent on the content of different
mances in both objective and subjective image quality.
images being processed.
In addition, three well-known 8-bit gray-level video se-
quences, Football, Akiyo and Erik, are used to test various
interpolation methods. The size of each frame is 352 × 288,

Table 1 Comparison of PSNR for different methods.

FOI [2] ELA [3] ELC [4] AMP [5] MBI [6] Ours
Baboon 23.42 22.90 23.10 23.46 23.40 23.44
Barbara 30.68 24.78 29.56 29.14 30.42 30.45
(a) Dd1 (b) Dd2 Couple 27.96 27.44 27.64 27.88 27.87 27.94
Crowd 31.76 31.19 31.37 31.70 31.68 31.82
F18 23.60 23.18 23.20 23.61 23.59 24.97
Indian 32.05 30.88 31.48 31.85 31.97 32.11
Lake 28.46 27.99 28.09 28.39 28.42 28.47
Lax 26.93 25.91 26.41 26.93 26.91 27.04
Lena 32.28 31.52 32.00 32.29 32.31 32.38
(c) Dv (d) Dh Peppers 29.22 29.33 29.36 29.42 29.36 29.36
Fig. 1 Four pixel differences for (a) Dd1 , (b) Dd2 , (c) Dv and (d) Dh . AVG 28.64 27.51 28.22 28.47 28.59 28.80

(a) (b) (c)

(d) (e) (f)

Fig. 2 Interpolated images of Lena for (a) Original, (b) FOI, (c) ELC, (d) APM, (e) MBI, and (f) Our
algorithms.
 IEICE TRANS. INF. & SYST., VOL.E90–D, NO.2 FEBRUARY 2007
608

Table 2 Comparison of average PSNR for three video sequences.

FOI [2] ELA [3] ELC [4] AMP [5] MBI [6] Ours 4. Conslusions
Football 35.99 34.65 34.51 35.68 35.93 36.16
Akiyo 39.94 37.90 39.46 39.55 39.95 40.24 A low-complexity spatial interpolation technique is pre-
Erik 29.94 29.43 29.71 30.01 29.95 30.06 sented to detect edges in the diagonal, horizontal and ver-
tical directions efficiently. The experimental results demon-
Table 3 Comparison of the worst-case PSNR for three video sequences. strate that our method achieves excellent performances in
FOI [2] ELA [3] ELC [4] AMP [5] MBI [6] Ours both objective and subjective image quality. Currently, the
Football 34.11 32.28 33.37 33.50 34.07 34.30 VLSI architecture of our method is under development.
Akiyo 39.64 37.32 38.86 39.07 39.52 39.83
Erik 29.51 28.88 29.18 29.52 29.48 29.59 Acknowledgment

This research was supported in part by the National Science

Table 4 Comparison of operations required for different methods.
Council, Republic of China, under the Grant NSC-95-2221-
FOI [2] ELA [3] ELC [4] AMP [5] MBI [6] Ours E-006-504. Also, this work made use of Shared Facilities
Absolute 0 3 7 4 5 5 supported by the Program of Top 100 Universities Advance-
Addition 1 1 3 5 3 5 ment, Ministry of Education, Taiwan.
Subtraction 0 7 13 19 25 11
References

and the frame rate is 30 per second. Each individual frame
is down sample by a factor of 2 vertically, and then inter-
polated by using different approaches. Tables 2 and 3 show
the average and worst-case PSNR of the first 50 frames for
the video sequences respectively. The proposed method
consistently produces better performance. This demon-
strates the capability of our method in coping with video
sequences.
For subjective testing, the interpolated Lena images
using different interpolation methods are shown in Fig. 2.
A portion of each image is enlarged in order to evaluate the
visual quality. FOI produces noticeable artifacts in areas of
angled edges as shown Fig. 2 (b). Like those previous pro-
posed methods [4]–[6], our method shows visually pleasing
picture especially at image edges.
To explore the computational complexity, we list the
operations required to interpolate each missing pixel for
various interpolation methods in Table 4. Compared with
the newly interpolation methods [4]–[6], our method re-
quires lower computational complexity (5 absolute, 5 addi-
tion and 11 subtraction operations). Although the proposed
method needs more operations than both FOI and ELA do, it
achieves better objective image quality as shown in Tables 1,
2, and 3.
[1] G.D. Haan and E.B. Bellers, “Deinterlacing–an overview,” Proc.
IEEE, vol.86, no.9, pp.1839–1857, Sept. 1998.
[2] R.S. Prodan, “Multidimensional digital signal processing for televi-
sion scan conversion,” Phillips J. Res., vol.41, pp.576–603, 1986.
[3] T. Doyle, “Interlaced to sequential conversion for EDTV applica-
tions,” Proc. 2nd Int. Workshop on Signal Processing of HDTV,
pp.412–430, Feb. 1988.
[4] T. Chen, H. Wu, and Z.H. Yu, “Efficient de-interlacing algorithm
using edge-based line average interpolation,” Optical Engineering,
vol.39, no.8, pp.2101–2105, Aug. 2000.
[5] H.-C. Kim, B.-H. Kwon, and M.-R. Choi, “An image interpolator
with image improvement for LCD controller,” IEEE Trans. Consum.
Electron., vol.47, no.2, pp.263–271, May 2001.
[6] M.Q. Phu, P.E. Tischer, and H.R. Wu, “A median based inter-
polation algorithm for deinterlacing,” Proc. International Sympo-
sium on Intelligent Signal Processing and Communication Systems,
pp.390–397, 2004.
[7] D. Wang, A. Vincent, and P. Blanchfield, “Hybrid de-interlacing al-
gorithm based on motion vector reliability,” IEEE Trans. Circuits
Syst. Video Technol., vol.15, no.8, pp.1019–1025, Aug. 2005.
[8] X. Gao, J. Gu, and J. Li, “De-interlacing algorithms based on mo-
tion compensation,” IEEE Trans. Consum. Electron., vol.51, no.2,
pp.589–599, May 2005.
[9] S. Yang, Y.Y. Jung, Y.H. Lee, and R.H. Park, “Motion compensa-
tion assisted motion adaptive interlaced-to-progressive conversion,”
IEEE Trans. Circuits Syst. Video Technol., vol.14, no.9, pp.1138–
1148, Sept. 2004.
[10] O. Kwon, K. Sohn, and C. Lee, “Deinterlacing using directional in-
terpolation and motion compensation,” IEEE Trans. Consum. Elec-
tron., vol.49, no.1, pp.198–203, Feb. 2003.