Scale-space ridge detection with GPU acceleration

Imaging systems for computer vision play an important role in today´s world. Typical computer vision systems operate on large scale scenes, where objects are relatively far from the camera and the depth of field in which objects appear focussed is large. Close-range camera systems, on the other hand, typically have a narrow depth of field. World features outside this depth of field are blurred, and in applications where poor data may not be re-acquired, a technique is required to reliably extract information from these images. Discrete scale-space feature detection techniques provide methods to extract features from these images, but bring with them a significantly higher computational workload compared with classical edge and ridge detectors. This paper presents the results from implementation of a discrete scale-space ridge detector with graphics processing unit (GPU) acceleration. This feature detector has been applied to close-range images of grids printed on sheet metal surfaces, and a speedup of one to two orders of magnitude is seen over a CPU-based implementation of the same feature detector.