Parallelism Research For Image Feature Extraction Algorithm

Nowadays, we have entered a data-centric world. A huge amount of data are transferred and processed on the Internet every day. Among these data, multimedia data have become more and more popular. Thus, it becomes vitally important to find an efficient way to manage and retrieval these multimedia data. As two most widely-used image feature extraction algorithms, SIFT (Scale Invariant Feature Transform) and SURF (Speedup Up Robust Features) are the kernel algorithms in image/video content-based retrieval and filtering systems. However, the limited processing speed (less than3frames per second on an ordinary CPU) makes them impossible to be applied in many real-time applications. Therefore, how to improve the processing speed of these image feature extraction algorithms becomes one of research hotspots.The popularity of multi-core architecture provides a new opportunity to accelerate these image feature extraction algorithms. In this paper, we first systematically analyze the parallel characteristics in these image feature extraction algorithms. We find that imbalanced workload greatly limits the performance of prior research. To alleviate these limitations, we design and implement a new parallel algorithm, adaptive pipeline (AD-PIPE), for these image feature extraction algorithms. In our scheme, different computations are partitioned into different pipeline stages to overcome the constraints of imbalanced workload. And the adaptive pipeline means that the algorithm can dynamically adjust the thread number in different stages according to the workload of these stages. This algorithm is efficient and scalable. The experimental results show that it can achieve a speedup of16.88X and20.33X respectively for SIFT and SURF on a16-core machine. The processing speed is about30and52frames per second by using16-core machine, which can satisfy the real-time requirements. As the accuracy is similar, SURF is much fast than SIFT and can gain more speedup on multi-core platforms. Thus, we think SURF is a better choice for large-scale image retrieval systems.