Proposal for Remote Sensing Image Analysis with Parallel-Processing Approach Using GPU

Abstract

We propose a method of acceleration for remote-sensing image analyses with the use of graphics processing unit (GPU). We applied the proposed GPU parallel processing methods to both filtering and correlation processes. We observed that the GPU acceleration increased with the moving window size for the convolution filter, because the convolution filter does not use any calculation arrays in the GPU shared memory. Since the median filter uses a sorting array in the shared memory, the acceleration reached the max in window size 9, and then decreased. We also investigated GPU parallel processing for correlations in both spatial and frequency domains. The area correlation method is a process using moving windows in the spatial domain, and it is possible to speed it up similar to the filtering process. As an example of frequency domain correlation methods, we investigated whether we could accelerate the phase-only correlation (POC). We developed a method to avoid the capacity constraint of GPU shared memory. By processing within the correlation window line by line, GPU could accelerate the POC even for the larger correlation window size exceeding 64. Our investigation of the GPU accelerations for filtering and correlation processes thus revealed that the important points are the reduction of the access load to global memory and avoiding the constraints of the shared memory size in the GPU.