ABSTRACT
In this article, Fast Global K-Means (FGKM) for Synthetic
Aperture Radar (SAR) image change detection is
presented.On account of the time-consuming of FGKM
algorithm and the real-time demand, we present a Parallel
Fast Global K-Means (P-FGKM) algorithm. We parallelize
the selection of initial cluster centers which is the most time-
consuming step of FGKM algorithm. The proposed
algorithm is implemented based on Open Computing
Language (OpenCL). The experiments are carried out on a
variety of heterogeneous computing devices , such as Multi-
core CPU, GPU, Intel HD Graphics, Many Integrated Core
(MIC). Experiment results show that the proposed algorithm
can achieve a good speedup up to 86 times on such devices.
Index Terms—Synthetic Aperture Radar, Image
Change Detection, Parallel algorithm , OpenCL, GPU
1. INTRODUCTION
In the past few years, the remote sensing image change
detection has become a research hot topic. It has a wide
variety of applications, such as civil infrastructure,
deforestation monitoring, underwater sensing, military
reconnaissance, disaster evaluation and video
surveillance[1].
Recently, a variety of methods to unsupervised SAR
image change detection are proposed. For example, Change
detection in SAR images by artificial immune multi-
objective clustering[2]. One of the most popular fuzzy
clustering algorithms is fuzzy c-means clustering (FCM)[3],
which partitions a given dataset
into c
fuzzy subsets.In addition, approaches based on clustering
algorithm which is a method of unsupervised learning that
partitions a set of data objects into clusters has been widely
studied. Because of the high performance of clustering
algorithm, relevant researchers have proceeded deep
research and presented many new clustering approaches to
the multi-temporal SAR image change detection in recent
years, such as approach using fuzzy clustering[4] and
method using FLICM algorithm[5] .
K-Means algorithm is a traditional local search
algorithm and one of the most popular clustering algorithms,
which suffers from the serious drawback that its
performance heavily depends on the initial cluster centers.
In order to solve the problem, A.Likas et al. proposed the
GKM algorithm in [6], a deterministic incremental
clustering algorithm. GKM algorithm comes from K-Means
algorithm, but the difference is that GKM employs a
deterministic approach rather than stochastic generation
method to obtain the initial cluster centers and does not
depend on any initial parameter values. However, GKM
algorithm has heavy computational cost, which is inefficient
to solve clustering problems for huge datasets. Thus the
FGKM [7] algorithm which accelerates GKM algorithm in a
straightforward method without significantly affecting the
quality of the results is proposed. Although FGKM is an
improved algorithm from GKM algorithm, it still has a huge
amount of computational cost.
Recently, with the development of science and
technology, mass data have appeared in every field.To deal
with huge amounts of data,a variety of heterogeneous
computing devices are used, such as Multi-core CPU, GPU,
Intel HD Graphics and MIC. Accordingly, to make use of
these heterogeneous computing devices, there are some
application programming interfaces. For example, OpenMP
can be applied to the CPU, while CUDA can be applied
to the NVIDIA GPUs. However, OpenCL has a good
portability, can be used on CPU, GPU,MIC,etc.
Parallel remote sensing image analysis, which has been
extensively studied in image processing [8], is still a
complex problem. However, the analysis of the
performance-portability of OpenCL on remote sensing field
is rare to see. In this article, we propose the SAR image
change detection based on FGKM algorithm. Due to the
time-consuming of FGKM algorithm and the high real-time
demand, parallelizing FGKM algorithm is a promising way
to overcome the challenge of the huge computational
requirement. In this article, our work includes the
parallelization of FGKM algorithm and the application of
FGKM algorithm to unsupervised SAR image change
detection. Furthermore, we verify the portability of P-
FGKM algorithm on heterogeneous computing devices,
such as multi-core CPU, GPU, Intel HD Graphics and MIC.