high performance computing on graphics processing units: hgpu.org

Scene recognition has become a remarkable field in image processing area, and many methods have been proposed in recent years, in which the idea of extracting the scene gist from global features has been proved to have higher retrieval accuracy compared with many other methods. However, the process of extracting gist is heavily time-consuming and not suitable for real-time application. In this paper, the CUDA architecture is deployed to accelerate this process. The influences of data transfer and memory access pattern on performance have also been investigated. Furthermore, when all the gist descriptors are got, parallel image feature similarity comparison is implemented with the OpenMP architecture. Experiments illustrate obvious speedup compared with implementation using only CPU.