Optimization of SIFT algorithm for fast-image feature extraction in line-scanning ophthalmoscope

The Scale Invariant Feature Transform (SIFT) algorithm is utilized broadly in image registration to improve image qualities. However, the algorithm's complexity reduces its efficiency in biology study and usually requires real-time. In this article, we present an improved SIFT technique in software architecture for matching sequences of images taken from a line-scanning ophthalmoscope (LSO). The method generates the Gaussian Scale-space pyramid in frequency domain to complete the SIFT feature detector more quickly. A novel SIFT descriptor invariable with rotation and illumination is then created to reduce calculation time, implementing the original SIFT method, our improved SIFT method, and the graphic processing unit (GPU) version of our improved SIFT method. The experiments have shown that the improved SIFT is almost 2-3 times faster than the original while maintaining more robust performance, and the GPU implementation of the improved SIFT is 20 times faster than central processing unit (CPU) implementation and achieves acceleration at real-time as expected. Although tested on an LSO system, the improved SIFT method does not rely on the acquisition setup. As a result, this method can be applied to other imaging instruments, e.g., adaptive optics to increase their resolution in agreement.