Abstract: This paper proposed a speed-up robust feature image registration algorithm based on Compute Unified Device Architecture (CUDA)．We analyzed the parallelism of SURF algorithm,and optimized it by CUDA from thread mapping and memory model of the Graphics Processing Unit (GPU),aiming at the five steps of SURF algorithm including building scale space,extracting feature points,determining key direction of feature points,generating vector descriptor and feature matching．The experimental results show that the GPU implementation achieves 33times faster than the CPU implementation while processing an image of 30MB．The GPU implementation extends the application of SURF algorithm in fast processing of remote sensing image,especially in the quick image registration.

Key words: Speed-up robust feature,CUDA,Feature extraction,Image matching