431
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
The static holds two important factors, one is the difference
between the gray value of current detection kernel and its local
neighbour pixels, and another is the difference between the
current detection kernel and the image background. Using the
two factors, we can detect the so called “anomalies”, whose
gray values are obviously lower or higher than the image
background.
When all the local clusters of anomalies are located, we find all
the positions of potential targets. According to the image
coordinate relationship between original images and resample
images, we can unprotect the detected clusters on the resample
images to the original images to locate some small image
blocks, in which one (or maybe a few) potential target should
be contained each.
2.3 Recognition
The output of Location step is a series of small located blocks
in original image containing one (maybe more than one when
two or three ships anchor together) potential target each. This
paper studies on two issues: is there ship in the block and what
type the ship is. Because there is not a universal automatic
algorithm for detecting unknown object, the effect of the
recognition result depends on the given prior knowledge.
Unsupervised classification (clustering) is a traditional method
to filter unexpected objects out in multi-spectral data, but it’s
not perfect for our detection task because sometimes other
objects show similar spectral characteristic. So a improved
regional segmentation algorithm based both on the geometry
features and spectral characteristic of ship objects is
conducted[Xu Da-qi,2006].In order to recognize the target fast,
knowledge about the shape and size of ship is required. The
segmentation algorithm is an iteration of well-known Otsu
segmentation, with an improvement in enhancing contrast of
gray value of the image.
3. EXPERIMENTS
To validate this method proposed in this paper, an aerial image,
with lm spatial resolution and RGB three spectral bands, which
is shot by digital frame camera in Yantai, Shangdong is used.
The original image size is 2616(width) by 3214(height). On the
sea surface there are 14 ships, four are navigating, the others are
anchoring and three ships are anchoring together like one ship.
The distribution of these ships is as following:
In searching step, a 3 X 3 kernel is used and the threshold of the
S value is set to 0.9.The searching results demonstrate the
efficiency of this method. By using the LISA analysis searching
method, all the anomalies are detected on the sampled image,
which performance as red spot in Figure 2.The whole searching
process on sampled image takes 0.133 second.
Figure 2.Searching results on the sampled image, whose size is
130(width) by 160(height).
A mathematical morphology dilation operation is processed on
the searching result to make anomalies that belong to the same
ship to connect to each other. Then each connected anomaly is
unprojected to the original image and located use red block,
which is a little larger than the anomalies area, shown in Figure
3. To avoid false alarm, a constrain is added to the detection
method that each ship block should have more than 3 anomaly
spots. So three blocks with only one spot in it are dropped. The
figure 4 shows the final detection result. All 14 ships are
detected, but in two blocks, ships anchoring together are
considered to be one ship by this method.
Figure 3. The located ships on original image
Figure 1. The original image.