Integral-image based implementation of U-SURF algorithm for embedded super parallel processor

In recent year, robust and scale invariant feature extraction algorithms such as SIFT, SURF, and U-SURF are frequently utilized for image recognition. While U-SURF algorithm is robust and scalable to extract interest points and their features, it requires processes in many regions that are scattered in an image. It is difficult for the algorithm to increase parallelism when an embedded parallel processor is used. In this paper, a parallelization method of U-SURF algorithm is proposed for implementation on a massively parallel embedded processor MX core. To realize the efficient algorithm, an integral image based parallelization method is utilized for feature vector extraction. From the result of performance evaluation with proposed implementation method, the MX core realizes the processing speed 40.6 times faster than that of the conventional RISC processor with same operating frequency. And also proposed feature extraction method realizes 7.3 times more efficient than original method. From these results, it is shown that the proposed algorithm is very effective for local feature extraction.