CC BY
11КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
Original article / Оригинальная статья УДК 504.054: 543.399
RESEARCH ON BAYER IMAGE INTERPOLATION OF SATELLITE VIDEO
© Aigong Xu , Jiaqi Wu , A.L. Okhotin
*School of Geomatics, Liaoning Technical University, Fuxin, Liaoning Province, 123000, China. **Irkutsk National Research Technical University, 83, Lermontov St., Irkutsk, 664074, Russian Federation.
ABSTRACT. In older to reduce the pressure of the satellite video data transmission, a satellite video captures the video data by using the Bayer pattern. To obtain the full color image sequences, Bayer image interpolation is required. Aiming at the interpolation of Video satellite Bayer Image, a method combining adaptive edge weights and the correlation of Inter-channel was developed. 12 Kodak standard images were used to make accuracy verification, the average sum PSNR value was more than 40. Using satellite video Bayer images to conduct reconstruction experiment, we compared the results of a developed method against some other classical algorithms'. Our method can perform a good work on the balance of edges' sharpness, color quality, noise suppression and zipper reduction. And it could be suitable for interpolation of satellite video Bayer images. Keywords: satellite video, Bayer pattern, Bayer image interpolation, edge weight
ИССЛЕДОВАНИЕ В ОБЛАСТИ ИНТЕРПОЛЯЦИИ ИМИДЖА БАЙЕРА ИЗ СПУТНИКОВОГО ВИДЕО Сюй Айгун*, Дзячи У*, А.Л. Охотин**
*Школа Геоматики, Ляонинский технический университет, Фусинь, провинция Ляонин, 123000, Китайакая Народная Республика. **Иркутский национальный исследовательский технический университет, Российская Федерация, 664074, г. Иркутск, ул. Лермонтова, 83.
For citation: Aigong Xu, Jiaqi Wu, Anatolii Okhotin. Research in Bayer Image Interpolation of a Satellite Video // XXI century. Technosphere Safety, 2017, vol. 2, no. 1, pp. 11-37. (In English).
Introduction
Satellite video is a new type of earth observation satellite. Comparing with the traditional satellite, the most salient feature of satellite video is that it can obtain more dynamic information by the mode of staring at one ground point and video recording. So, it specializes in observing the dynamic objects and analyzing transient characteristic. If the change rate of continuous frames surpasses 24 frame per second, the visual effect will be
smoothness based on THE CYCLOTROPE, and this is called a video. Currently, the frame rate of a satellite video has not been limited confirm internationally. It is true for the geostationary orbit diffraction imaging system of the United States, which can shoot a video with 1 frame per second rate. The European geostationary orbit satellite space surveillance system (GO-3S, which is being demonstrated, can shoot a video with 5 frame per second
Aigong Xu, e-mail: ag@126.com Сюй Айгун, e-mail: ag@126.com Jiaqi Wu, e-mail: ag@126.com Дзячи У, e-mail: ag@126.com
**Anatoly L. Okhotin, Candidate of technical sciences, Professor, Head of the Department of Mine Surveying and Geodesy, e-mail: ohotin@istu.edu
Охотин Анатолий Леонтьевич, кандидат технических наук, профессор, зав. кафедрой маркшейдерского дела и геодезии, e-mail: ohotin@istu.edu
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
rate. The sky satellite 1 skysat-1) launched by the United States can obtain video with 30 frame per second rate [1].
Gazing refers to the optical imaging system stare at a target region with the satellite motion, and it can observe the change continuously in the view field. There are two ways of "Gazing", one is a geostationary orbit optical imaging satellite, the other is a low-orbit optical imaging satellite able to posture agile or image motion compensation. The geostationary orbit satellite is stationary relative to the ground because of the orbital dynamics. So, it can implement "Gazing". But if the image resolution reaches the meter level, the caliber of an imaging system should be big enough. At present, the United States and Europe are actively developing a large-caliber (at least greater than 4m) optical imaging system. The "Gazing" low-orbit satellite can be divided into two classes. One class is the satellite provided with linear-array detector with the ability of high agility, represented by the US WorldView and France Pleiades. The other one is the satellite provided with an area array detector implement "Gazing" with the platform high agility ability. The typical instances are LAPAN-Tubsat developed by Indonesia and Germany, and skysat-1 launched by the United States.
A video satellite can form a video by using a sequence image with a time interval which is suitable for analyzing dynamic objects, obtaining the velocity and direction of the objects. This dynamic information can hardly be obtained from the traditional static image.
Internationally, the geostationary orbit optical imaging satellite with a meter-level resolution, which is actively developing, have the ability of video recording for a long time. The satellite is widely used in the domain of marine and environment surveillance, and could detect the large military moving targets [2].
A single low-orbit video satellite can only observe a target region persistently for about 1
min on account of the high speed over the top. So, a single satellite could not monitor targets operationally and persistently. The small size video satellite developed by foreign scientists adopt the microsatellite platform, the quality is about 100 kg. So the low cost can support satellite constellation disposition, and implement the object surveillance near-real time. The foreign low-orbit video satellite has been changed from an experiment research to operational application. The resolution has been developed from 5 m to 1 m. So, it can implement the fast motion detection, confirmation and recognition with the cooperation of low-orbit sub-meter resolution optical imaging satellite [3].
In order to obtain the full color image, we need at least 3 color gray at each pixel location. A method is performing the RGB color separation, and extracting each color by employing 3 detectors. This method can obtain high quality images, but it will make the HD satellite video a great amount of data which is bad for data transmission and storage. So, in order to reduce the pressure of the satellite video data transmission, researchers employ only one detector in the camera imaging system. And a color filter array (CFA) is installed on the surface of the detector, the commonly used CFA is Bayer pattern. The image produced by a Bayer pattern called a Bayer image which is used in this paper. The Bayer pattern sample graph is shown in Fig. 1. When shooting image, we get only one color gray in each pixel location, and the other two colors are filtered. The b, c, d in Fig. 1 are red, green and blue channels of a Bayer image. A Bayer image reduces two thirds of the data, which can save resources of the satellite system. Nevertheless, a Bayer image lost color information, we must perform the interpolation reconstruction processing firstly, after that, the full color image obtained can be used in the further application.
12
м^и
ISNN 2500-1582
Vs
Я Ш
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
G R G R
В G В G
G R G R
В G В G
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
Fig. 1. Bayer's pattern Рис. 1. Образец Байера
H
d
The basic method of Bayer image interpolation is divided into following types: linear interpolation, inter-band correlation, edge gradient, iterative and frequency method et al. Linear interpolation [4] mainly includes neighborhood interpolation and bilinear interpolation, which perform an interpolation in the single channel and have a fast-operating rate. But the reconstruction quality is unsatisfactory. The inter-band correlation refers to a pixel change in each color band is basically coincide, that means the color ratio [5] and color difference [6] between green component and red or blue component are low pass and change smoothly. This method produce false color artifacts and blurriness. Edge gradient is a method that extracts the gradient image using the local pixel, and determining the edge direction employing the local gradient. After that, we preform the linear interpolation along the edge direction [7, 8]. The edge gradient method can make a good interpolation effect, but the difficult points are edge judgment accurately. The iterative method obtains the full color image by adopting the linear interpolation in a general way. Afterwards we used the iterative function designed to approximate the raw color image step by step [9]. The iterative method can produce a good quality color image, but the complexity is generally high, and convergence often hard to guarantee. The iterative times should be appointed. The frequency method exploits the signal correlation to reconstruct the full color image by using the bandpass filter in the frequency domain, and the method has high efficiency. But it is easy
to produce zipper [10].
In view of the edge gradient method with high precision, many scholars make a deeper study based on the method. DLMMSE [11] exploits the local minimum MSE principle to perform the interpolation. The neighborhood information has been effectively used and the precision of results is high, but it costs much time. Chung [12] determines the edge by using the neighborhood color difference, and adopt different interpolation depending on whether the missing pixel lies on the edge. The effect is excellent, but the time cost is also much. Su performs the favorable effect edge interpolation with edge weight based on the color difference. The high and low frequency region are distinguished by using gradient, and adopt edge gradient method in the high frequency region [13, 14]. Zhu Bo [15] perform the gradient direction interpolation in the color difference domain after Laplacian correction. The last 3 method all make a good effect, but they do not consider the variation of weight and correlation. MSG [16] determine the edge direction by using the multi-scale gradient, and perform the adaptive interpolation based on the color difference estimation. This method could not set threshold value, but cost a lot of time. DB [17] determine the interpolation direction according the correlation of neighborhood and inter-band, and exploit the nonlocal self-correlation to reduce the false color artifacts, but it also cost much time.
In a full-color image, the gradient contains important detailed information [18, 19]. The edge direction of the pixel can be obtained by
H
Том 2, № 1 2017 Vol. 2, no. 1 2017
XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
f3f
vOy
b
a
c
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
using the gradient. Adjacent pixels in the edge direction have higher gray relativity, and interpolating along the direction gives better results. But the non-edge direction of the adjacent pixel gray correlation is not high, but also contains a certain color information. If the image resolution is high, the non-edge direction of the correlation will increase. In addition, the correlation of the edge gradient is not the same at different positions, the amount of color information contained is variable: The larger the gradient, the higher the corresponding correlation. The edge direction should be given a
greater weight, on the contrary, smaller. Classical Bayer interpolation algorithm generally only consider the edge of the high correlation or use the fixed weight to calculate the missing pixels of gray, which is clearly not accurate enough. Therefore, considering the dynamic correlation between pixel neighborhoods, an adaptive edge weighting method is proposed, which can extract the edge weights efficiently, and combine the interpolation of gradient and inter-band correlation to reconstruct the Bayer image. The effect and accuracy are both well.
2. Description of the Developed Algorithm
The green component pixel of the Bayer image lacks 1/4, more gray scale and the frequency information are reserved, the green band is preferentially interpolated, and then the red and blue bands are interpolated based on the green band. The general idea of the proposed method is as follow. For the green band, first using the neighborhood and adjacent bands to obtain the horizontal and vertical gradient, and according to the gradient to determine the edge weights and interpola-
tion direction. And then we perform the gradient interpolation with the weight along the edge direction. After the green band interpolation, the color ratio domain of G/R and G/B can be established, then interpolating the missing pixels with equal weight in the color ratio domain, and then converting to the corresponding band to complete the red and blue image reconstruction. The overall flow of the method is shown in Fig. 2.
Fig. 2. Algorithm flow chart Рис. 2. Блок-схема алгоритма
14
М^И
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
2.1. Determination of the Edge
Weight
As shown in Fig. 3, R G B represents the red, green and blue pixels at that location respectively. Taking the R7 position as an example, the G-grayscale is estimated, taking into account the correlation between adjacent bands, and the gradient is corrected using red band pixels based on the Laplacian principle. Horizontal gradient SH is:
SH = \G6 - G8| +12 * R7 - R5 - R9\ + Sh , (1)
Fig. 3. The interpolation pattern Рис. 3. Образец интерполяции
Sh is a very small value, which avoid the horizontal gradient is zero. Likewise, the vertical gradient SV is:
SV = |G3 - G11| +12 * R7 - R1 - R13|. (2)
Horizontal and vertical gradient corresponding weight cH and cv are:
(3)
СОн+Юу= 1
2.2. Interpolation Reconstruction of a Green Band
First, the horizontal and vertical gradations of the pixel to be interpolated are calculated. The gray-scale value is modified by us-
ing the adjacent band and the neighborhood pixels based on the Laplacian. Horizontal direction of the gray is:
Gh = 1/2* ( G6 + GS + R7 )--1/4 ( RS + R9) .
(4)
Similarly, the vertical direction of the
gray is
Gv = 1/2* ( G3 + G11 + R7 )--1/4 ( R1 + R13)
(5)
The direction gray level contains the color information of the pixel to be interpolated and occupies a certain proportion according to the gradient correlation of the pixel and the area in which it is located. When the difference between the horizontal and vertical gradient is large, the pixel is located at the edge region, and the edge direction is determined according to the numerical comparison of the two-direction gradient. The corresponding weight is given to the gray of different directions, and the adaptive direction interpolation is performed. When the difference between the horizontal and vertical gradient is small, it means that the pixel is in the smooth region, the weight of the gradient can be considered equal, and the equal-weight interpolation is performed. The calculation formula is:
\G = Gh * cH + Gv * С if \SH-SV\ > Th |G = 1/2* (Gh + Gv) else
(6)
In the above formula, Th is the gradient difference threshold. Through the adaptive edge weight direction interpolation, it not only can adapt to the edge of the sensitivity, but also can reduce the smooth region of the zipper effect. For the interpolation of missing G positioned on the blue, we take the same above.
H
Том 2, № 1 2017 Vol. 2, no. 1 2017
XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
15
2
4
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
2.3. Blue (red) Color Band Interpolation Reconstruction
After green band interpolation reconstruction, the color ratio space of G/B and G/R is established. In the color ratio space, the whole image is smooth and the high frequency region is less [5]. Therefore, interpolation in the color ratio space can obtain good interpolating effect by using equal weight, and guarantee the efficiency. The Laplacian is also used to obtain the gradient. The three positions G3, G6 and R7 in Figure 3 are the three cases of blue interpolation (horizontal, vertical, and neighborhood). For G3 position of B, only horizontal B2 and B4 are known, that is the interpolation direction. So, we perform the linear interpolation directly in the color ratio space.
B3 = G 3/2* ( B2/G'2 + B4/G'4). (7)
Where G'2, G'4 correspond to G value of B2, B4 position, obtained by interpolation above. The B value of the G6 position is directly interpolated by B2 and B10 in color ratio
space, as same as B3.
For B interpolation of position R7, the interpolation direction is judged by horizontal and vertical gradients. The color ratio horizontal gradient is:
ÔHBG = |B'6/G6 - B'8/G8|. (8)
Color ratio vertical gradient is:
ÔVBG = |B'3/G3 - B11/G11|. (9)
The direction of the smaller gradient is interpolation direction, for equal direction interpolation is:
I B'6/G б-. , B = l/2*G 7*I if SHRr <SV„r
+B'S/GS 1 BG BG
iB'3/G3 + . B = l/2*G 7*I else.
+B'll/ Gll,
(10)
The red band interpolation is the same
as blue.
3. Evaluation and Experiment
There are two kinds of evaluation method of Bayer interpolation reconstruction, which are subjective evaluation and objective evaluation. First, subjective evaluation is based on the subjective visual perception of the observer to make judgments and evaluation of the image. Second, objective evaluation, calculating the deviation between the interpolated images and truth image by using some statistical calculation method. Commonly used statistical methods are mean square error (MSE) and peak signal to noise ratio (PSNR). The MSE calculation formula is:
2
1 m |2
MSE =—YU1 (x, y) - y) (11)
m*n j j 1 1
MSE can directly express the deviation between interpolation image and the original image in the pixel gray value, and describe the accuracy of the image interpolation. The smaller MSE is, which indicate the accuracy of Bayer interpolation is high, the closer to the real. But it cannot evaluate the objective quality of reconstructed images. While PSNR, based on the MSE, is the most widely used evaluation of the objective quality of the Bayer interpolation measurement method. Calculation formula is:
(2b -1)2
PSNR = 10 • log((-—4 (12)
MSE
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
H
Where b is the color depth of the image. Kodak standard images will be adopted for numerical evaluation of image interpolation and measuring the interpolation accuracy of various algorithms. On this basis, the satellite video Bayer images are used for visual evaluation of interpolation, so as to measure the interpolation effect of various algorithms. Numerical and visual evaluation can perform comprehensive evaluation for the method proposed.
3.1. PSNR Numerical Statistics of Kodak Images
The Kodak standard image shields the optical system differences of the different imaging devices and the signal errors caused by the sensors and their corresponding hardware. In the field of CFA interpolation algorithm research, researchers usually use Kodak standard color images for quality evaluation. 12 Kodak standard image sets are selected as truth images, as shown in Fig. 4. Sorting from left to right, top to bottom, the selected image avoids people and too densely populated textures to approximate image features to satellite remote sensing images. The experimental steps are: firstly, the Bayer images of selected truth images are respectively generated according to the Bayer pattern. And then, the interpolated image was reconstructed by our algorithm. Finally, the numerical statistic is performed for the accuracy of these full color images interpo-
lated. The numerical statistic of the proposed algorithm is shown in Table.
PSNR can not only describe the deviation between the interpolation image and the truth image, but also can evaluate the quality of the image objectively. The larger the PSNR value, the better the image quality and the less distortion. In general, if the PSNr is greater than 40, the interpolated reconstructed image quality is very good and very close to the truth image. When the PSNR is between 30 and 40, the image quality is good and the distortion is detectable but acceptable. According to the average value of the results in Tab., red and blue bands of the proposed method is very close to 40, and the green band almost reaches 43, interpolation image quality and precision are high. However, the PSNR are sometimes not consistent with the visual quality observed by the human eye. Because the human eye to the sensitivity of the error is relative, for example, some interpolated images have high PSNR values, and only have obvious color false in some specific texture-rich regions. However, due to the high sensitivity of the human eye to color false, the subjective image quality is considered to be poor. Therefore, based on the objective quality evaluation of interpolated image, the interpolation effect needs to be further observed by the human eye, and do more depth and comprehensive evaluation.
H
Fig. 4. Kodak images Рис. 4. Изображения, полученные с помощью Кодака
Том 2, № 1 2017 Vol. 2, no. 1 2017
XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
шж 1l
к^к/
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
PSNR value of a developed method
Ценность PSNR предложенного метода
Image PSNR
R G B
2 SS.S2 42.40 S7.SS
4 S4.S9 S9.47 3б.26
б 41.1S 44.6s 42.42
б SS.27 42.4б S9.71
7 41.4S 4б.б3 41.42
S S9.12 42.S7 40.SS
9 41.1S 44.10 S9.12
10 S9.19 42.S9 S7.91
11 S7.4S 40.12 S7.0S
17 S7.SS 41.4S SS.0S
1S S9.66 42.SS 40.S2
22 40.S4 4б. 11 41.б7
Avg S9.11 42.7S S9.S0
3.2. Comparison and Analysis of a Bayer Interpolation Experiment for Satellite Video images
For the satellite video images, we perform the experiment of Bayer interpolation reconstruction. The selected data was a city in Mexico, with a resolution of 1.1m. The typical features were selected for comparison, including motor vehicles, highways and houses. In order to facilitate the details of observation, we intercept the interest regions which magnified several times, and perform the linear stretch treatment. The results comparison between the proposed algorithm and several classical algorithms as shown in Fig. 5, 6. Several classical algorithms selected are: TSD [14], DLMMSE [11], DB [17] and MSG [16]. Test environment: MATLAB R2015b.
Observe the reconstructed results of Bayer images in Fig. 5, 6. The TSD uses the comparison of horizontal and vertical gradients to determine the direction of interpolation and the division of high and low frequency regions. Reconstruction quality is good, the edges are sharp and clear. But because of using a fixed weight to judge edge, and using red-green and blue-green difference for the correlation correction. So, that some of the edges near the
smooth area, especially the areas with high red and blue gray levels, the grainy and zipper effect are obvious. Such as the roof of the house and the vicinity of road divider in Fig. 6, these areas are unsatisfactory. DLMMSE uses the method of linear least mean square error to estimate the missing pixels, and the reconstruction quality is good. The distortion of reconstruction image is not significant, and the color is true. But this method assumes that the signal of the green-red and green-blue bands is low-pass. This is not completely true in the edge region. So, the missing pixel grey estimation of the edge area is not accurate enough. The buildings and road edges in Fig. 6 are slightly blurred, the contrast between vehicle and road in Fig. 5 are slightly lower. DB also uses the inter-band correlation to determine the direction of interpolation. While, the local self-similarity principle is used to improve the interpolation precision in the high-frequency region. The edge of the reconstructed image is sharp and clear. But when the low contrast between object and surrounding environment, or noise interference happens, self-similarity will regard this as correlation of neighborhoods, leading to interpolation errors. In Fig. 6 the smooth region noise is ob-
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
vious, in Fig. 5 the vehicle color saturation is slightly lower, especially the color of motor vehicle is close to the road, the interpolation effect is not good. MSG uses a multi-scale gradient method, can accurately determine the neighborhood correlation and interpolation direction, which can ensure image reconstruction quality. The feature edges are clear, transition of color is natural. As shown in Fig. 5,
the car contour is obvious with high identifiabil-ity. But there is also sparse noise in the results, and this method is time consuming. The method proposed has high color saturation, color true and natural transition. The edges of the features are sharp and clear. Especially, a contrast ratio of vehicle on the road is relatively high, obvious and clear. And there is less zipper effect and noise.
rS¡.
\Y
Vf
d e f
Fig. 5. Comparison of results of RO11: a - Bayer; b - TSD; c - DLMMSE; d, e, f - MSG, ours
H
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
ÏS
vöy
b
a
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
Fig. 6. Comparison of results of ROI 2: a - Bayer; b - TSD; c - DLMMSE; d, e, f - MSG, ours
rn^rn
Том 2, № 1 2017 Vol. 2, no. 1 2017
XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
Vs
Я Ш
b
a
c
e
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
4. Conclusion
In this paper, we studied the Bayer interpolation method of a satellite video. Two key points of interpolation are interpolation direction and direction correlation were dealt with. For the key points, the interpolation direction is confirmed accurately by the dynamic edge weights and the correlation between the bands, and the direction correlation is quantized based on the weight and direction gray levels. High and low frequency region interpolations are distinguished by a gradient differ-
ence, while using the smoothness of color ratio to improve the accuracy of red and blue bands. Experimental results show that the developed algorithm has good interpolated reconstruction results and a high numerical precision. Comparing with the classical algorithm, the comprehensive effect of our method is better and it could be suitable for the full color reconstruction of the video satellite Bayer image.
Acknowledgements
'This study was supported by the Chi- (GDW20162100099), and the Liaoning Prov-na National Key R&D Program ince University Innovation Team Project (2016YFC0803102, the 2016 Liaoning Prov- (LT2015013). ince Top Foreign Experts Project
References
1. Liu Tao. Research on the Development of Foreign Video Satellite [J]. Space international, 2014, no. 9, pp. 50-56.
2. Zhu Lihong, Hui Zheng, Ren Defeng, [et al]. Development Status and Inspiration of Video Imaging Satellite [J]. Satellite applications, 2015, no. 10, pp. 20-28.
3. Blu T., Thévenaz P., Unser M. Linear interpolation revitalized [J]. IEEE Trans: Image Process, 2004, vol. 5, no. 13, pp. 710-719.
4. ZHANG Xiangfei, ZHANG Gang, CHENG Yongqiang. The Designing and Realization of High resolution Bayer CFA Interpolation Algorithm Based on FPGA [J]. JOURNAL OF TAIYUAN UNIVERSIT Y OF TECHNO-LO GY, 2006, no. 37, pp. 12-15.
5. Lu W, Tan Y.P. Color filter array demosaicking: new method and performance measures [J]. IEEE Transactions on Image Processing, 2003, vol. 10, no 12, pp. 1194-1210.
6. Lukac R., Martin K., Plataniotis K.N. Demosaicked image post-processing using local color ratios [J]. IEEE Trans: Circuits and Systems for Video Technology, 2004, vol. 6, no. 14, pp. 914-920.
7. Yang Huizhen, Cheng Yongqiang. A low Computational Complexity Demosaicing Method for Bayer Pattern CCD [J]. JOURNAL OF TAIYUAN UNIVERSIT Y OF TECHNOLO GY, 2007, vol. 3, no. 38, pp. 215-218.
8. Niruban R., Arthi M., Sree Dr.T., Raja R. Directional color difference based demosaicing [J]. International
Journal of Advanced Research in Electronics, Communication & Instrumentation Engineering and Development, 2014, vol. 1, no. 2, pp. 19-28.
9. Gunturk B., Altunbasak Yucel, Mersereau Russell M. Color plane interpolation using alternating projections [J]. IEEE Transactions on Image Processing, 2002, vol. 9, no. 11, pp. 997-1013.
10. Jeon G., Dubois E. Demosaicking of noisy bayersampled color images with least-squares luma-chroma demultiplexing and noise level estimation [J]. IEEE Transactions on Image Processing, 2013, vol. 1, no. 22, pp. 146-156.
11. Zhang L., Wu X. Color demosaicking via directional linear minimum mean square-error estimation [J]. IEEE Transactions on Image Processing, 2005,14(1): 2167-2178.
12. Chung K H, Chan Y H. An edge-directed demosa-icing algorithm based on integrated gradient[C]. IEEE International Conference, Multimedia and Expo (ICME), 2010, pp. 388-393.
13. Su C.Y. Highly Effective Iterative Demosaicing using weighted-edge and color-difference interpolation [J]. IEEE Transactions on Consumer Electronics, 2006, vol. 2, no. 52, pp. 639-645.
14. Su C.Y. Low-complexity hybrid demosaicing for color filter arrays [J]. Journal of the Chinese Institute of Engineerings, 2008, vol. 1, no. 31, pp. 173-180.
15. ZHU Bo, WEN Desheng, WANG Fei. Improved
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
КОНТРОЛЬ И МОНИТОРИНГ ОПАСНОСТЕЙ HARARD MANAGEMENT AND MONITORING
Bayer pattern demosaicking and its hardware design [J]. Journal of Optoelectronics Laser, 2013, vol. 6, no. 24, pp. 1211-1218.
16. Poornila A., Kingsta M. Multiscale Gradients based directional demosaicing [J]. International Journal for Trends in Engineering & Technology, 2015, vol. 1, no. 4, pp. 90-98.
17. Duran J., Buades A. Self-similarity and spectral correlation adaptive algorithm for color demosaicking [J]. IEEE Transactions on Image Processing, 2014, vol. 9, no. 23, pp. 4031-4040.
Authorship criteria
Aigong Xu, Jiaqi Wu, Anatolii Okhotin have equal authors rights and responsibility for plagiarism.
Conflict of interests
The authors declare no conflict of interests.
Received on 16.02.2017
18. ZHU Bo, WEN Desheng, WANG Fei, [et al]. Improvement of Bayer-pattern Demosaicking with Dictionary Learning Algorithm [J]. Journal of Electronics & Information Technology, 2013, vol. 4, no. 35, pp. 812-819.
19. Kim Y., Jeong J. Subdivided weight interpolation based on multiscale gradients for color filter array [C]. Proceedings of the 2014 International Conference on Communications, Signal Processing and Computers, 2014, pp. 32-35.
Критерий авторства
Сюй Айгун, Дзячи У, Анатолий Охотин обладают равными авторскими правами и несут равную ответственность за плагиат.
Конфликт интересов
Авторы заявляют об отсутствии конфликта интересов.
Поступила 16.02.2017
22
М^И
Том 2, № 1 2017 XXI ВЕК. ТЕХНОСФЕРНАЯ БЕЗОПАСНОСТЬ Vol. 2, no. 1 2017 XXI CENTURY. TECHNOSPHERE SAFETY
ISNN 2500-1582
