An enhanced SURF algorithm based on new interest point detection procedure and fast computation technique

In this paper, we propose an enhanced Speeded Up Robust Features (eSURF) algorithm to save memory and increase the operating speed. From analysis and observation of the conventional SURF algorithm, we show that a large amount of memory is inefficiently used to detect interest points and considerable operations are repeatedly performed when generating the descriptors of interest points. In the proposed algorithm, the scale-space representation (SSR) step and location (LOC) step are unified based on an efficient memory allocation technique to remove unnecessary memory. In addition, operations for Haar wavelet responses (HWRs) in horizontal and vertical directions, which occupy a major portion of computational loads, are performed by using a fast computation technique in which redundant calculations and repeated memory accesses are efficiently eliminated. Simulation results demonstrate that the proposed eSURF algorithm achieves a time savings of approximately 30% and a memory savings of approximately 35.7%, while the feature extraction performance of the proposed eSURF algorithm is exactly identical to that of the conventional SURF algorithm.