Efficient implementation of the COSFIRE algorithm with CUDA

Lovdal, Sofie (2018) Efficient implementation of the COSFIRE algorithm with CUDA. Bachelor's Thesis, Computing Science.

Preview	Text Thesis_Sofie_v4.0.pdf Download (3MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (91kB)

Abstract

The thesis explores runtime performance improvement of the COSFIRE algorithm when parallelized with CUDA on GPUs. COSFIRE is a trainable filter for pattern recognition in images. The core algorithm consists of four steps which are applied a number of times with different parameters, creating several subresponses. The steps involve convolution with a Difference-of-Gaussians filter, max-blurring and shifting. The subresponses are gathered by geometric mean. In this algorithm both task parallelism and data parallelism can be identified. The CUDA implementation exploits the massively parallel architecture of a GPU by employing mainly data parallelism - in each step of the algorithm, one thread performs the calculations for one pixel. While the control flow of the CUDA implementation is kept in host code, the amount of data transfer between the CPU and the GPU is kept to a minimum and all key steps of the algorithm are executed on the GPU. On a GTX 1080 ti GPU with 3584 cores, a speedup of 28.3 has been achieved when applying the CUDA COSFIRE implementation on a retinal fundus image of resolution 999x960 pixels.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Biehl, M.
Degree programme:	Computing Science
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	27 Jul 2018
Last Modified:	06 Sep 2018 12:21
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/18092

Actions (login required)

View Item