MORPHOLOGY INFRARED IMAGE TARGET DETECTION ALGORITHM OPTIMIZED
BY GENETIC THEORY
a State Key Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan
University, P.R. China, -shaozhenfeng - 163.com, xianqiangzhu@163.com
b School of Remote Sensing and Information Engineering, -rsliujun@mail.whu.edu.cn
KEY WORDS: Genetic Algorithm, Morphology Filter, Optimization Multi-Group Evolution, Target Detection, Infrared Image
ABSTRACT:
This paper proposes a novel morphology algorithm for target detection of the infrared images, which is optimized by the genetic al
gorithm. First, several improvements have been adopted for genetic algorithm (GA). The improvements include the auto-regulation
of genetic evolution crossover probability and mutation probability which based on population difference; A more reasonable target
characteristic variable according the feature of infrared image has been designed to train structure elements. So the efficiency of the
algorithm can be improved obviously, at same time, the optimized trained structural elements could reflect the sample’s true structure
information. Then the background and objectives structural elements could be got by inputting background and target samples into
the GA model. Using it as a priori knowledge of the morphology operation, it does favor to improve the algorithm’s accuracy and
adaptability. .Experiment shows that this method can achieve a higher detection efficiency and accuracy.
1. INTRODUCTION
Infrared has been widely used in the target detection technology
for its special recognition capability and the anti-jamming
ability. However, due to the detection range of the infrared
detectors and other conditions, infrared images are
characterized by severe noise pollution and low signal-to-noise
ratio. Therefore there is a growing interest in the research of the
methods which are used in the target detection of the infrared
images. In recent years, mathematical morphology which based
on shape and structure information has been paid more and
more attention in the area of target detection for its superior
performance in the field of infrared target detection, however,
the results of the target detection are due to the choice of the
structural information mentioned above.
The idea of the genetic algorithms (GAs) which derived from
the theory of evolution is different from the traditional
optimization methods. GAs can be used to solve complex
unstructured problems for its global optimal search capability
and parallel computing characteristic. Using GAs, we can get
the optimal structure elements which contain samples’
information by training the Genetic Evolution Model (GEM)
with only a small sample of the origin images. This paper
demonstrates how multi-populations GA can be employed in
the target detection using improved mathematical morphology.
IMPROVED GENETIC ALGORITHM
GA has particular advantages in solving the problem of multi
parameter optimization for its global search mechanism, but on
contrary to traditional GA there are also some shortcomings that
can not be ignored, such as the powerful global search capabil
ity while poor local search ability , premature problem, the con
tradictions between the search efficiency and accuracy. Aiming
at improving these weaknesses, many scholars have done a lot
of work to improve its performance, such as the introduction of
population difference, improved genetic operators and so on. In
this paper, the genetic algorithms is used into the field of target
detection and a new fitness measurement methods based on tar
get characteristics value is proposed for the specific needs of
detection. For the premature problem, basing on the population
difference which is proposed by predecessors, the introduction
of the male and female parents dynamic clustering based on in
dividual differences makes crossover operate more reasonable;
The strategy of multi-population parallel evolution and best in
dividual transplant merge the best individual retention strategy
and the best gene retention strategy effectively, which acceler
ate the speed of convergence of the algorithm and meanwhile
ensure the global search capability.
Population Difference
From the perspective natural evolution, if the group owns a
great diversity larger crossover and smaller mutation ratio is the
best choice. On the contrary, smaller cross scale and larger mu
tation probability is better choice, so we can see population dif
ference is a useful variable in determining evolution parameters.
Details refer to the references [1].
2.2 Encoding Method
In this paper, we have chosen a real-parameter coding
method which avoids information loss during the encoding and
decoding process. Compared to binary coding method, every
individual in the population contains original information of the
structure elements. In the abstract this coding method has a
more appropriate complexity and higher accuracy result.
Group Initialization