CUDA架构下高效红外图像背景预测方法

Efficient infrared image background prediction in CUDA

红外图像背景预测方法是红外图像弱小目标探测与跟踪中的经典方法.针对背景预测算法中卷积运算耗时长的问题,提出一种基于CUDA(Compute Unified Device Architecture)的高效红外图像背景预测方法.在分析背景预测算法执行流程的基础上,充分考虑CUDA架构的特点,将其在CUDA架构下进行了重新实现,利用GPU(Graphic Processing Unit)的强大并行计算能力完成红外图像背景预测的快速计算.为了进一步提升算法的运行效率,将不可分离的背景预测卷积模板分解为多个可分离模板的叠加,并给出了分解卷积模板的一般方法.将该方法应用于实际的红外图像背景预测,结果表明,该方法比传统的CPU计算在运算效率上提高了130倍以上.

Background prediction in infrared is a classical method for small and dim target detection.Taking account of the convolution＇s time-consuming property,an efficient method for infrared image background prediction based on CUDA（Compute Unified Device Architecture） is proposed.By considering the feature of CUDA and analyzing the process of background prediction,it is re-realized.Using the framework of CUDA and taking advantage of the powerful parallel computing ability of GPU（Graphic Processing Unit）,background prediction from infrared images is rapidly and efficiently completed.For further raising the efficiency of the method,the inseparable convolution matrix used by background prediction is broken up into several separable matrices,where a general method for separating the convolution matrix is presented.Applying this method to actual infrared image＇s background prediction,the computation efficiency is raised more than 130 times compared with the traditional method based on CPU.