The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
1238 
the result will store in the texture m_fbortl[l] no matter at 
which level. 
Be the same, the FBO of the second image binds two textures 
too: m_fbort2[0] and m_fbort2[l]. The result of the 2D DWT 
will be stored in the texture m_fbort2[l]. 
The next step is to fuse the data in texture m_fbortl[l] and 
m_fbort2[l]. The texture m_fbortl[l] and m_fbort2[l] are bind 
and passed to weighted fusion fragment program by two 
uniform variables the type of which is samplerRECT, and the 
destination buffer is set GLCOLORATTACHMENTOEXT. 
Then pass the data in this buffer to 
GLCOLORATTACHMENT1EXT. The invert DWT is a 
reconstruction process from the highest level. The result data of 
the reconstruction at level j is the source data of level j-1. As the 
case of forward 2D DWT, the result of 2D inverse DWT is 
stored in buffer GL COLOR ATTACHMENT1 EXT. At last 
download the data in texture m_fbortl[l] bind to 
GL COLOR ATTACHMENTI EXT to system memory to get 
the fusion result. 
Color Image Fusion Algorithm Based on IHS and Wavelet 
Transformation, APPLICATION RESEARCH OF 
COMPUTERS, 23(10), PP. 174-176. 
[4] Tien-Tsin Wong., Chi-Sing Leung, Pheng-Ann Heng, 
Jianqing Wang,2004. Discrete Wavelet Transform on 
Consumer-Level Graphics Hardware. ACM Workshop on 
General-Purpose Computing on Graphics Processors (GP2). 
6. EXPERIMENT RESULT 
Experiment condition: DELL PC, 3GHz Pentium(R) 4, 
NVIDIA GeForce 8800 GTS with 320M video memory; 
Experiment data: Landsat TM images. We choose 4 groups of 
registered images and the size (pixel) of each group is: 
256x256, 512x512, 1024x1024, 2048x2048. 
Table 1 shows the time cost of the fusion arithmetic both on 
CPU and GPU. 
Image size 
256x256 
512x512 
1024x1024 
2048x2048 
product 
fusion 
GPU 
8.0 
16.8 
51.0 
184.4 
CPU 
16.5 
65.4 
263.9 
1098.3 
ratio 
fusion 
GPU 
8.1 
17.0 
50.9 
184.1 
CPU 
8.1 
33.1 
136.9 
533.5 
high-pass 
filtering 
GPU 
10.3 
17.1 
45.0 
150.0 
CPU 
3.6 
14.5 
59.4 
297.9 
weighted 
fusion 
GPU 
7.6 
16.4 
50.4 
181.8 
CPU 
6.6 
26.8 
106.9 
454.4 
IHS 
fusion 
GPU 
60.3 
89.4 
195.5 
636.8 
CPU 
42.8 
172.1 
650.2 
2621.4 
DWT 
fusion 
GPU 
647.5 
770.7 
1043.1 
2135.5 
CPU 
419.3 
1627.8 
5808.1 
23910.5 
Table 1. Time cost comparision of GPU and CPU 
From the result we can see that the image fusion based on GPU 
do not has the advantage as based on CPU while the data size is 
small, but with the image size getting bigger, the image fusion 
speed based on GPU is much quicker than that based on CPU, 
and the computing time on GPU does not increase 
proportionally with the data size, and the advantage increases 
with the complexity of the fusion arithmetic. 
REFERENCES 
[1] Randi J.Rost,2006. OpenGL Shading Language. Posts & 
Telecom Press, Beijing, pp.83-96 
[2] Jia Yonghong,2005. Multi-sensors Remote Sensing Image 
Data Fusion. Surveying and Mapping Press, Beijing, pp.31-38. 3 
[3] TANG Guo-liang, PU Jie-xln , HUANG Xin-han,2006.