GPU-based SIFT algorithm
|School||Huazhong University of Science and Technology|
|Course||Spatial Information Science and Technology|
|Keywords||Image Registration Heterogeneous computing SIFT GPU|
Image registration is an important image processing technology for many more advanced image processing technology provides basic support. Depending on the selected mode, the image registration method is mainly divided into two categories : area-based ( gray ) of the registration and feature-based registration . Feature-based registration due to its performance and stability, flexible approach , wide availability, has been widely studied and applied. SIFT algorithm is currently the more popular a stable feature-based registration method , because of its scale and rotation invariance of light and perspective transformation also has better robustness , becoming academic research and engineering applications hotspots . SIFT algorithm , while having good performance , but its lack of speed and efficiency , limiting him in more fields of application, such as video retrieval, target tracking, object recognition time-sensitive applications . To solve this problem , in order to improve the SIFT 's performance in these applications , the paper introduces a GPU CPU heterogeneous computing platform , we propose a heterogeneous SIFT algorithm implementation , the SIFT algorithm more time-consuming scale space generation, Gaussian convolution arithmetic -intensive operations to the GPU to achieve full use of ultra- large-scale streaming GPU processor array and the advantages of high-speed on-chip memory to improve the efficiency and speed of SIFT algorithm , to lay the foundation for further research and application . Feature-based image registration methods are generally process is divided into four steps: feature extraction, feature space generated , find matching features , image registration . This article will focus on the extraction of this step . In Microsoft Visual Studio 2008 Express, OpenCV 2.1, CUDA 3.2.16 integrated platform , this study achieved on SIFT algorithm for heterogeneous experiments were developed and constructed a heterogeneous system image barebones . The final result proved , GPU SIFT algorithm can effectively improve the speed of the registration , GPU CPU heterogeneous platforms capable of compute-intensive image registration algorithm brings speed improvement.