面向互联网应用的图像LBP算法GPU并行加速

GPU-based acceleration of local binary pattern algorithm for Internet applications

很多互联网应用需要进行海量的图片处理。LBP算法是一种图像纹理特征提取算法,广泛用于图像检索等领域,但该算法较为复杂,在处理互联网环境中的海量图片时面临着性能挑战。解决该问题的办法之一就是采用GPU对LBP算法进行并行加速,特别是针对海量图片处理设计加速方案,使GPU同时进行多幅图像LBP特征的并行提取,并采用异步传输方式使多幅图像数据的复制与Kernel函数的执行并行化。通过对GPU单幅和多幅图像并行处理的实验测试,并将实验数据与CPU程序性能进行对比分析,结果表明:对不同分辨率多幅图像并行处理的加速比可达58倍。

Many Internet applications need to process massive images. As a typical image texture fea- ture extraction algorithm, Local Binary Pattern （LBP） is widely used in image retrieval and other fields. This algorithm is relatively complex, so it faces the performance challenges on processing massive ima- ges on the Internet. One solution to this problem is using the Graphics Processing Unit （GPU） to sup- port parallelization of extracting Local Binary Pattern texture feature, especially an accelerating scheme for the processing of massive images. The GPU can compute the LBP feature of multiple images in par- allel. And asynchronous transmission is used to perform the data replication and kernel function of mul- tiple images in parallel. The parallel processing of single image and multiple images using the GPU are tested, and the performance of the program on the GPU is compared with the one on the CPU. It is shown that the parallel processing for different resolution multiple images can achieve 58 times speedup.