Professional Documents
Culture Documents
Abstract This paper describes an object detection system The PCNN-based iterative image smoothing technique
based on pulse coupled neural networks. The system is de- is shown to smooth images without blurring, eroding, or
signed and implemented to illustrate the power, flexibility and dilating edges. The thin lines and curves present in the
potential the pulse coupled neural networks have in real-time
image processing. In the preprocessing stage, a pulse coupled image that are often completely removed by the median
neural network suppresses noise by smoothing the input image. filter degrade gracefully in the new approach [3]. The
In the segmentation stage, a second pulse coupled neural-network PCNN based segmentation method, under certain conditions,
iteratively segments the input image. During each iteration, guarantees perfect segmentation of the input image even
with the help of a control module, the segmentation network when the intensity ranges of adjacent regions overlap.
deletes regions that do not satisfy the retention criteria from
further processing and produces an improved segmentation of the Perfect segmentation conditions are derived in [1]. The
retained image. In the final stage each group of connected regions inclusion of inhibition receptive field in the neuron model
that satisfies the detection criteria is identified as an instance of significantly improves the segmentation capabilities of the
the object of interest. PCNN by compressing the intensity ranges of the individual
Index TermsImage smoothing, image segmentation, iterative regions in the image and by reducing the extent of overlap
segmentation, object detection, pulse coupled neural network, of the intensity ranges of adjacent regions. Simulation
pulse coupled neuron. results indicate that the PCNN yields better segmentation
results than most of the commonly used methods [2].
I. INTRODUCTION
Fig. 3. Smoothed image of the tank. The tank in Fig. 2 was smoothed with
= 0:02, r = 1:5 and C = 2 over 4 pulsing cycles.
Fig. 1. Object detection system.
and the neurons corresponding to the pixels of adjacent regions of the segments elongation. For a given area, radius of
do not pulse together. gyration assumes its minimum value when the segment
Assume that the image to be segmented consists of is circular. Its value is maximum when all pixels are
regions and is applied as input to a PCNN. The network collinear. A segment whose radius of gyration exceeds
neurons pulse based on their feeding and linking inputs. Note the expected maximum value of the radius of gyration of
that the feeding input to a neuron is equal to the intensity of the tank cannot represent the tank or a part of the tank.
its corresponding pixel. Due to the capture phenomenon the The maximum value of the radius of gyration of a single
neurons associated with each group of spatially connected pix- segment (which is same as the radius of gyration of the
els with similar intensities tend to pulse together. Thus, each tank) was determined to be 14 pixels in length.
contiguous set of synchronously pulsing neurons identifies a In order to eliminate a segment from further consideration
segment of the image. The value of the linking coefficient the feeding inputs (pixel intensities) of the corresponding
has significant effect on the segmentation process. If the neurons are set to a unique negative value. This will also allow
value of is low, the pixels that belong to a single region are the control module to enable the previously disabled segments
partitioned by the PCNN into several segments. This is known selectively.
as over-segmentation. On the other hand, if the value of is If the number of active segments at the end of the current
high, the PCNN groups pixels that belong to two or more iteration is less than a prespecified value (25 segments), the
regions to form a single segment. This is known as under- segmentation process is terminated. The resulting segments
segmentation. It may be possible to find the optimal value are input to the detection module for further processing. Oth-
of based on the intensity probability density function of erwise, the value of is increased by a constant amount and
the image and the geometry of objects present in the image. the process of segmentation (Step 2) and segment elimination
However, the determination and the use of the optimal value (Step 3) is repeated.
of does not guarantee perfect results. The justification for gradually increasing the value of
Since the determination of the optimal values of parameters starting from a low initial value is given in this paragraph.
is not always possible, an iterative approach is used to segment It is assumed that there is some contrast between the regions
the image. In this approach, values of all parameters except of tank and the surrounding background regions. If this is not
remain constant throughout the object detection process. true the tank cannot be distinguished from the background by
The PCNN begins the first iteration with a low value of any method. When the contrast between the tank and its sur-
to ensure over-segmentation. Its value is increased slightly by rounding regions is low the risk of the tank or a part of the tank
the control module for each subsequent iteration. The rationale pulsing with the background increases. Therefore, a low initial
for choosing this approach is described later. At the end of a value of is chosen to ensure that the tank segments do not
pulsing cycle all the segments are presented to the control merge with the background segments due to increased linking
module. activity. During the early iterations, any segment that is larger
Step 3: Image segments to be discarded are determined. than a tank represents homogeneous background regions such
The knowledge base includes information to facilitate the as sky and meadow. Similarly, large and elongated regions
evaluation of individual segments and the merging of adjacent represent features such as roads and rivers. Thus, they may be
segments. The actual size and shape of the tank in the image eliminated from further consideration. After eliminating large
varies depending on the orientation, the imaging system and homogeneous regions, the image is fragmented and consists of
the imaging geometry. Model based algorithms and special many less homogeneous regions. The value of is increased
purpose hardware are available for the real-time estimation of to reduce fragmentation.
the values of these attributes for any desired object. Therefore, In the first iteration, the segmentation of the image in
it is decided to store in the knowledge base 1) the maximum Fig. 3 with and yielded 61 segments
and minimum sizes of the tank; 2) the maximum value of the as shown in Fig. 4(a). Six regions that did not satisfy the
radius of gyration of segments; and 3) the maximum distance size or elongation constraints were deleted. The black regions
between the centroids of adjacent segments that are candidates in Fig. 4(b) represent the deleted regions. Note that more
for merging. The control module examines each segment and than 50 percent of the pixels were eliminated from further
discards segments that violate any of the following constraints. consideration. Since the number of segments was greater than
Area constraint: The upper and lower limits of the area 25 the image in Fig. 4(b) was segmented again. The value
of the tank in the image are determined by the imaging of was increased by 0.03 at the end of each iteration. The
system parameters and the distance between the tank and results of five iterations are summarized in Table I. Note that
the camera. If the area of a segment exceeds the upper the number of segments decreases as the processing continues
limit, it is unlikely to represent a tank. Therefore, such from one iteration to the next. During the fifth iteration only
large segments are eliminated from further consideration. one segment was deleted to yield a total of 21 segments as
For the set of images used in simulation, the lower and shown in Fig. 4(c). Therefore, the segmentation process was
upper limits for the area of the tank were determined to terminated.
be 480 and 700 pixels, respectively. Step 4: Detection of a group of connected regions that could
Elongation constraint: In addition to its area, the elonga- form an object of interest.
tion (spatial extent) of a segment is an important property. Each segment received by the detection module may be
The radius of gyration of a segment is a good measure a tank or a part of a tank. The function of the detection
618 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 10, NO. 3, MAY 1999
determining optimal value for the linking coefficient. The FL, JuneJuly 1994, pp. 12851290.
use of very simple knowledge has succeeded in reducing the [2] G. Kuntimad, Pulse coupled neural network for image processing,
Ph.D. dissertation, Dept. Comput. Sci., Univ. Alabama, Huntsville,
number of regions that need to be processed by the recognition 1995.
system. The use of additional knowledge is likely to improve [3] H. S. Ranganath, G. Kuntimad, and J. L. Johnson, Pulse coupled neural
the performance further. networks for image processing, in Proc. IEEE Southeastcon, Raleigh,
NC, Mar. 1995.
[4] M. K. Hu, Visual pattern recognition by moment invariants, IRE Trans.
REFERENCES Inform. Theory, vol. IT-8, pp. 179187, 1962.
[5] H. S. Ranganath and G. Kuntimad, Iterative segmentation using pulse
[1] H. S. Ranganath and G. Kuntimad, Image segmentation using pulse coupled neural networks, Applicat. Sci. Artificial Neural Networks II,
coupled neural networks, in Proc. Int. Conf. Neural Networks, Orlando, SPIE Aerosense, Orlando, FL, vol. 2760, pp. 543554, Apr. 1994.