Accelerating Image Feature Comparisons using CUDA on Commodity Hardware

Given multiple images of the same scene, image registration is the process of determining the correct transformation to bring the images into a common coordinate system—i.e., how the images fit together. Featurebased registration applies a transformation function to the input images before performing the correlation step. The result of that transformation, also called feature extraction, is a list of significant points in the images, and the registration process will attempt to correlate these points, rather than directly comparing the input images. The Scale-Invariant Feature Transform [1], or SIFT, is a popular feature extraction algorithm. After finding significant points, it generates a descriptor for each point. These descriptors are a collection of circular histograms describing the intensity gradients of small regions surrounding the point. The structure of these descriptors, combined with multi-scale extraction, makes SIFT feature descriptors invariant to the rotation and scale of the input images. Georeferencing, which locates images in physical space, is one application of image registration. The location of an input image, for example an aerial photo, can be determined by registering it against other aerial images with known coordinates. When using featurebased registration, this is implemented by extracting features from the input image, searching for similar features in the existing images, and calculating a coordinate transformation that correlates the similar features. Searching for similar features implies the existence of a function to measure feature dissimilarity. One such function is the circular earth mover’s distance [2], or CEMD, which provides excellent feature comparison but requires more computation than other methods. Georeferencing frequently operates on large images, requiring searches among billions of features, so a method to quickly perform many CEMD comparisons is prerequisite to the use of CEMD in georeferencing applications.