Implementation And Speedup Of Contourlet-based Image Fusion Algorithm Based On Da Vinci Platform

As the image fusion technology is widely used in medical image, military industry and spaceflight, research on the image fusion becomes more and more popular. The classic wavelet transform has been applied in image fusion for decades. Ever since Do and Vetterli proposed a new theory namely, contourlet transform indicated its advantages with better efficiency. Contourlet-based image fusion algorithms emerged. However, the even higher computational complexity it requires turns out to be a disastrous concern in embedded applications.In this paper, we implement an acceleration system on a heterogeneous TI Da Vinci dual core processor consisting of an ARM processor core and a DSP processor core. First, we give introduction and analysis of the image fusion theory and image transforms. We analyze the features of wavelet-based image fusion and contourlet-based image fusion through examples of images. Then we give comparison between wavelet and contourlet through comparing the quantitative indexes including standard deviation, average grads and entropy, proving that contourlet transform has better performance than wavelet. We implement the contourlet-based image fusion on Da Vinci platform. On ARM core only, the whole program consumes 67 seconds. In order to accelerate the image fusion, we dispatch different tasks to different processor. We put the I/O tasks and control tasks to ARM core and matrix computation tasks to DSP core. Through this way, the whole program only needs 8 seconds, with speedup 8.4. At last, we analyze the experiment results. We design test case and get that the speedup of DSP core on matrix computation is 36.91. We can use Amdahl Law to get the totally speedup, which proves our experiment results. We also prove that the overhead of ARM core calling DSP core can be ignored for comparing to the total time consumed, the overhead is too small to be considered.