Abstract
Distance transform (DT) and Voronoi diagrams (VDs) have found many applications in image analysis. Euclidean distance transform (EDT) can generate forms that do not vary with the rotation, because it is radially symmetrical, which is a desirable characteristic in distance transform applications. Recently, parallel architectures have been very accessible and, particularly, GPU-based architectures are very promising due to their high performance, low power consumption and affordable prices. In this paper, a new parallel algorithm is proposed for the computation of a Euclidean distance map and Voronoi diagram of a binary image that mixes CUDA multi-thread parallel image processing with a raster propagation of distance information over small fragments of the image. The basic idea is to exploit the throughput and the latency in each level of memory in the NVIDIA GPU; the image is set in the global memory, and can be accessed via texture memory, and we divide the problem into blocks of threads. For each block we copy a portion of the image and each thread applies a raster scan-based algorithm to a tile of m×m pixels. Experiment results exhibit that our proposed GPU algorithm can improve the efficiency of the Euclidean distance transform in most cases, obtaining speedup factors that even reach 3.193.
Highlights
The distance transform is an operator applied, in general, to binary images, composed of foreground and background pixels
We present a new parallel algorithm for the computation of the Euclidean distance map of a binary image
For 1024 × 1024 pixel resolution images, the Parallel Banding Algorithm (PBA)+ algorithm performs better on 1% density images, but our Parallel Raster Scan for Euclidean Distance Transform (PRSEDT) algorithm achieves better results for images with 10% to 90% density, with speedup factors ranging from 1.763 to 3.717
Summary
The distance transform is an operator applied, in general, to binary images, composed of foreground and background pixels. The result is an image, called a distance map, with the same size of the input image, but where foreground pixels are assigned a numeric value to show the distance to the closest background pixel, according to a given metrics. The distance map for an input image of n × n pixels can be computed, obviously, in O n4 time using brute force. Depending on the employed metrics, distance information from the neighborhood of a pixel can be exploited to compute the distance value, and different algorithms have been proposed in order to exploit the local information [2]. The algorithms can be classified in parallel, sequential, and propagation approaches
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
