CPU/GPU Cooperative Fast Band Registration Method for Multispectral Imagery

With the rapid increase of data size of remote sensing images, the traditional serial band re-gistration method cannot meet the demand for real-time processing of big-data multispectral images. Therefore, a CPU/GPU cooperative fast band registration method for multispectral imagery is proposed in this paper. Firstly, the computational amount and degree of parallelism are analyzed; point matching and differential rectification are ported to GPU to execute while the affine transformation parameter is still calculated on CPU. Secondly, kernel task assignment and basic settings are made to ensure the two above GPU steps executable. Moreover, three performance optimization methods, including memory access optimization, instruction optimization and transmission/computation overlap, are designed to further improve the efficiency of band registration. The experimental results based on NVIDIA Tesla M2050 GPU and Intel Xeon E5650 CPU show that the running time of YG-26 multispectral image band registration is only 3.25 s with our method, which got a speedup ratio of 32.32 compared with the traditional CPU serial method. The proposed method can provide quasi-real-time processing capability for multispectral imagery with big data size.