Efficient SIFT processing using sub-sampled convolution and masking techniques

The Scale Invariant Feature Transform (SIFT) is an algorithm for describing local features in an image. This research successfully demonstrates a model for optimizing SIFT using block convolution pre-filtering. A method is presented which theoretically reduces SIFT run time by nearly 50% by greatly limiting the area of image regions required to search for SIFT features. The block convolution of an image is computed using a kernel designed to be predictive of potential regions of interest. The performance of several different types of kernels are compared, most of which were produced using an evolutionary search algorithm. The filtered image is used to locate the potential regions of interest and mask regions that are unlikely to produce matches. This model will provide the framework for an optimized hardware implementation of SIFT in scenarios requiring low power and high speed, such as a robotic computer vision system. The advantage of the model lies in the very low computational intensity required to mask areas of the image where a match is unlikely to be found, yielding a more efficient implementation of SIFT.