CDVS EXTRACTION PROCESS ACCELERATION METHOD BASED ON GPGPU PLATFORM

摘要

Disclosed is a CDVS extraction process acceleration method based on a GPGPU platform, comprising two steps: feature detection and local descriptor calculation of a CDVS extraction process. The arithmetic logic and parallel strategy of various parallelizable sub-phases between pixels thereof and various parallelizable sub-phases between feature points are realized by means of an OpenCL general parallel programming frame, and the acceleration of the parallel calculation capability of a GPU is realized. The method comprises: dividing calculation tasks of a GPU and a CPU; reconstructing a scale space pyramid storage model of an image; assigning a parallel strategy for various sub-phases on the GPU; and applying a local memory to weaken the memory/access bottleneck. By using the solution, the CDVS extraction process can be accelerated, and the extraction performance is improved.