Abstract
Synthetic image generation is a key process in a change-based disaster detection method which detects differences before and after a disaster from Synthetic Aperture Radar (SAR) images using a convolutional neural network. In this process, a synthetic image representing the situation just before the disaster is generated by the following three steps: (1) histogram matching, (2) convolution, and (3) similar pixel search; among them, the latter two steps are time-consuming. In this paper, we propose a GPU acceleration method for the synthetic image generation, mainly for steps (2) and (3). The step (2) is parallelized in a patch-by-patch basis to apply a convolution operation to each patch in each SAR image. While the step (3) is parallelized in a pixel-by-pixel basis to search the most similar pixel on the corresponding position in multiple SAR images. Due to the limitation of memory capacity on a GPU, each SAR image is divided into several regions to be searched at a time. The evaluation results show that the reduction rate of the execution time for steps (2) and (3) is about 89.3% and 98.5%, respectively, compared with Python execution on a host computer.