Speeded-Up Image Feature Extraction Algorithms On GPU

Currently, compared to traditional text data, video and image data has become one of the major data types transferred and processed in Internet. However, many applications of images and videos, such as search engines and network information filtering systems are far from applicable due to the constraint of image retrieval algorithm. There are mainly two classes of image retrieval algorithms:one is based on global features and the other is based on local features. Retrieval algorithms based on global features describe an image or a frame of a video with a unique feature of color, texture, shape or space relationship. They run fast but can’t satisfy image retrieval’s need because of their low precision. On the other hand, retrieval algorithms based on local features describe an image or a frame of a video with hundreds of features and have high precision. But they also can’t be used because image feature extraction algorithms based on local features are much too complicated and run slow. So, the key point is how to accelerate image feature extraction algorithms based on local features.In the last few years, with the development of semiconductor and popularization of multi-core technology, parallel hardware has become the mainstream of applications. GPU has become an integral part of it because it’s becoming more and more general-purpose and programmable. The modern GPU is not only a powerful graphic processor but also a highly parallel programmable processor. Stronger arithmetic ability and higher bandwidth bring GPU great competitive power in real-time processing systems and high-performance computing systems.This paper designs and implements parallel acceleration algorithms for SIFT and SURF on GPU. We make best use of GPU architectures to accelerate our implementations, including uses of shared memory and texture memory, decreasing allocation and free times of memories on GPU and etc. GPU usually cooperates with CPU to finish a job. But traditional optimization strategies pay too much attention to the implementation on GPU and ignore CPU, which can be carefully used to achieve higher performance of the whole system. This paper implements SIFT and SURF on GPU in consideration of the impacts of CPU. Evaluations with640*480size images show that SIFT achieves the speedup of143.7X and the throughput of93.39frames per second and SURF achieves the speedup of253.2X and the throughput of346.82frames per second, which satisfies real-time processing’s need of image feature extraction. Considering the processing speed gap between SIFT and SURF, we recommend SURF as image feature extraction algorithm in real-time processing systems.