High resolution digital elevation models (DEM) used in digital terrain analysis are becoming more and more prevalent. There are various curvilinear structure extracting algorithms

but the main limitation of them is the high computing cost

making them less efficient when extracting topographic feature lines from high resolution DEM images. We propose an efficient strategy to speed them up using Steger's curvilinear structure detection algorithm imple mented on graphic processing units (GPU). We choose to speed up the most computation intensive modules of the algorithm (Hessian matrix generation and feature point detection) using NVIDIA's compute unified device architecture (CUDA). This method can achieve more than five times speedup compared with the original algorithm on central process units (CPU) for large scale DEM images with millions of pixels.