Turbopix算法的CUDA并行实现

CUDA-based parallel implementation of Turbopix algorithm

过分割是计算机视觉领域流行的图像预处理方法。针对其运行速度慢的缺点,对广泛采用的Turbopix算法提出CUDA并行优化的方法。通过每个线程执行一个超像素扩张的任务分配,实现了水平集函数的并行演化;利用纹理存储空间和常数存储空间的优化策略,改善了数据访存的效率。实验结果表明,在GT 240M平台上,平均加速比达到了15以上。

Oversegmentation is a popular method for image preprocessing in computer vision. In order to overcome its inten sive computational requirement, a CUDA based parallel implementation for widely used Turbopix algorithm is proposed. The parallel levelset evolution is implemented via superpixel expansion in each thread. The texture and constant memory optimization tech niques are also adopted to improve performance. Experimental results in a GT 240M graphics card show that a speed up of more than 15X on average is achieved.