Dir Ectional Bicubic Inter Polation - A New Method of Image Super - Resolution
Dir Ectional Bicubic Inter Polation - A New Method of Image Super - Resolution
Dir Ectional Bicubic Inter Polation - A New Method of Image Super - Resolution
1 Intr oduction
As high-resolution (HR) images can provide more details and better perceptual
quality, image interpolation is required to produce a HR image from its low-
resolution (LR) counterpart in many fields, such as medical images, HD videos,
satellite imaging, mobile devices (e.g. iphone, ipad, etc) and so on.
Before interpolation, the first step is to detect local edge direction. As gradients
across edges are larger than along edges, local gradients are used for detecting
edge directions. For a digital image, four directions (0°, 45°, 90°and 135°) are
considered in a 7×7 neighborhood as follows:
G0 = ∑ ∑ I (i + m, j − n ) − I (i + m, j − n + 2) + ∑ I (i + m, j − 1) − I (i + m, j + 1)
m = ±1n = 3, ±1 m = ±3
G90 = ∑ ∑ I (i − m, j + n ) − I (i − m + 2, j + n ) + ∑ I (i − 1, j + n ) − I (i + 1, j + n )
m = 3, ±1 n = ±1 n = ±3
471
G45 = ∑ I (i + 1, j − n ) − I (i − 1, j − n + 2 ) + ∑ I (i + 3, j − n ) − I (i + 1, j − n + 2)
n = 3, ±1 n = ±1
+ ∑ I (i − 1, j − n ) − I (i − 3, j − n + 2)
n = 3,1
∑ I (i + m, j − m ) − I (i + m − 2, j − m + 2)
1
+
2 m =3, −1
G135 = ∑ I (i − 1, j − n ) − I (i + 1, j − n + 2) + ∑ I (i − 3, j − n ) − I (i − 1, j − n + 2)
n = 3, ±1 n = ±1
+ ∑ I (i + 1, j − n ) − I (i + 3, j − n + 2)
n = 3,1
∑ I (i − m, j − m ) − I (i − m + 2, j − m + 2)
1
+
2 m =3, −1
(1)
Suppose the direction across edge is between the biggest two gradients
G max 1 , G max 2 of G0 , G 45 , G90 , G135 . Compute the ratio of G max 1 , Gmax 2 , where 1 is
added to avoid division by zero. If (1 + Gmax 1 ) (1 + Gmax 2 ) > T , pixel (i, j ) is on a
strong edge, the direction of the edge is across G max 1 ; otherwise, pixel (i, j ) is on a
weak edge, the direction of the edge is across a direction between the two
gradients.
In the paper, we divide pixels into two classes: pixels on strong edges and pixels
on weak edges or textures. Take the interpolation of pixel I HR (2i,2 j ) for example.
I HR (2i,2m )
472
1 9 9 1
− 32 I HR (2i −1, 2 j − 3) + 32 I HR (2i −1, 2 j −1) + 32 I HR (2i −1, 2 j +1) − 32 I HR (2i −1, 2 j + 3)
− 1 I +
9
I +
9
I −
1
I
32 HR (2i +1, 2 j − 3) 32 HR (2i +1, 2 j −1) 32 HR (2i +1, 2 j +1) 32 HR (2i +1, 2 j + 3)
θ = 0
1 9 9 1
− I HR (2i + 3, 2 j − 3) + I HR (2i +1, 2 j −1) + I HR (2i −1, 2 j +1) − I HR (2i − 3, 2 j + 3)
16 16 16 16
θ = 45
I HR (2i , 2 j ) = (2)
− 1 I HR 9 9 1
(2 i − 3, 2 j −1) + I HR (2i −1, 2 j −1) + I HR (2i +1, 2 j −1) − I HR (2i + 3, 2 j −1)
32 32 32 32
1 9 9 1
− I HR (2i − 3, 2 j +1) + I HR (2i −1, 2 j +1) + I HR (2i +1, 2 j +1) − I HR (2 i + 3, 2 j +1)
32 32 32 32
θ = 90
1 I 9 9 1
− 16 HR (2 i − 3, 2 j − 3) + 16 I HR (2 i −1, 2 j −1) + 16 I HR (2 i +1, 2 j +1) − 16 I HR (2 i + 3, 2 j + 3)
θ = 135
1
w1 = 1 + G k
max 2
(3)
w2 = 1
1 + G max
k
1
473
Interpolating pixels I HR (2i,2 j ) . Before interpolating, first compute its local
variance σ . If σ < Tσ , use bilinear interpolation directly. Otherwise:
Compute gradients of four directions as (1).
Extract the two biggest gradients G max 1 , Gmax 2 , and interpolate the pixel p as
follows:
In this section, we compare our results with four methods: Bicubic interpolation
[1], NEDI [4], PCAI [5] and DCCI [6]. Fig.2 displays the test images.
474
3.1 Experiment Results
Table 1 reports the PSNR results under five methods for two times zoom. It can be
observed that the proposed method has the best result, which is 0.46dB, 0.44dB,
0.42dB, 0.15dB higher than Bicubic [1], NEDI [4], PCAI [5] and DCCI [6]
respectively. Table 2 and Table 3 compare structural similarity (SSIM) [7] and
feature similarity (FSIM) [8] results, which also show that our method is better.
Table 1 Comparison of five different methods in PSNR (dB)
Images Bicubic[1] NEDI[4] PCAI[5] NCCI[6] Proposed
475
Fig.3 shows partition results of ‘Airplane’. It can be seen that the proposed
method has the best results. Bicubic interpolation [1] has the most severe blurring
(Fig. 3(b)). NEDI [4] breaks the geometric duality between the LR covariance and
HR covariance, and generates speckle noise (Fig. 3(c)). PCAI [5] simply divides
directions into four classes which doesn’t match the structure of real images, and
has obvious blurring artifacts (Fig. 3(d)). DCCI [6] generates shape edges but has
weak subjective results (Fig. 3(e)). Fig.4 and Fig.5 are partition results of
‘Airplane’ and ‘Doll’, which also show similar results.
476
Table 4 Average percentage of pixels interpolated by bilinear interpolation and the proposed
method
Percentage Bilinear Proposed
4 Conclusion
Refer ences
1. Keys, R. (1981). Cubic convolution interpolation for digital image processing. Acoustics,
Speech and Signal Processing, IEEE Transactions on, 29(6), 1153-1160.
2. Hou, H., & Andrews, H. (1978). Cubic splines for image interpolation and digital filtering.
Acoustics, Speech and Signal Processing, IEEE Transactions on, 26(6), 508-517.
3. Jensen, K., & Anastassiou, D. (1995). Subpixel edge localization and the interpolation of still
images. Image Processing, IEEE Transactions on, 4(3), 285-295.
4. Li, X., & Orchard, M. T. (2001). New edge-directed interpolation. Image Processing, IEEE
Transactions on, 10(10), 1521-1527.
5. Yang, B., Gao, Z., & Zhang, X. (2012, July). Principal Components Analysis-Based Edge-
Directed Image Interpolation. In Multimedia and Expo (ICME), 2012 IEEE International
Conference on (pp. 580-585). IEEE.
6. Zhou, D., Shen, X., & Dong, W. (2012). Image zooming using directional cubic convolution
interpolation. Image Processing, IET, 6(6), 627-634.
7. Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). Image quality assessment:
From error visibility to structural similarity. Image Processing, IEEE Transactions on, 13(4),
600-612.
8. Zhang, L., Zhang, L., Mou, X., & Zhang, D. (2011). FSIM: a feature similarity index for
image quality assessment. Image Processing, IEEE Transactions on, 20(8), 2378-2386.
477