Image Processing Reference
In-Depth Information
In order to simplify matters and also because the features present in a particular class are not
present in its super class, we analyze each building block separately in the following subsections.
However, the reader should keep in mind the existing relationships indicated in Figure .. Sections
.. through .. describe each category in greater detail and illustrate the different methods that
have been employed to incorporate the various features.
4.4.3 In-Network Processing
While the precise manner in which in-network processing is carried out on sensor nodes may differ
between various publications, the fundamental objective is still the same—to save energy by reducing
message transmissions. Directed diffusion [] performs in-network processing using filters. When a
node receives a message that matches its filter, the message is first handed over to the application
module within the node for processing instead of forwarding it to the next node. For example, the
application might carry out a suppression of duplicate messages indicating the detection of the occur-
rence of an event so as to prevent a sudden burst of identical messages when a bunch of nodes detect
thesameevent.hemaindrawbackofdirecteddifusionhowever,isthatithasanondeclarative
interface and thus does not rank very highly in terms of usability. As shown in Table ., directed
diffusion is designed primarily to extract aggregated data from a certain part of the network over an
extended period of time.
Cougar [], which was one of the first projects to view a sensor network as a distributed database
with a declarative interface, uses a clustered approach to in-network processing. A network may con-
sist of several clusters each of which is made up of a single leader node and a group of child sensor
nodes belonging to the leader. Child nodes periodically send their readings to the leader node which
then aggregates the received readings and only continues to forward the computed result toward
the root of the network. Computation at the leader only takes place once all the child nodes have
responded. Additionally, since sending multiple small packets is more expensive than sending one
larger packet (due to the packet header payload and the cost of reserving the medium) Cougar per-
forms packet merging by combining several packets into one. his method is particularly beneficial
when servicing queries that generate holistic aggregates where intermediate nodes cannot perform
any partial aggregation and all data must be brought together to be aggregated by the node evaluat-
ing the query, e.g., the Median operator or even the collection of raw sensor readings. It is important
to note, however, that while Cougar claims to be designed for WSNs, it was deployed on PDA-
class devices that had significantly larger processing power and could even run Windows CE and
Linux []. [].Their design does not consider the impoverished power and computational constraints
of conventional sensor nodes, e.g., XML, which is known for its verbosity, is used to encode messages.
Apart from packet merging, Cougar lacks any other features that make it energy-efficient. It also fails
to address the issues of reliability and self-organization and relies instead on the underlying .
MAC protocol.
TinyDB [] supports a number of aggregation operations (e.g., MIN, MAX, SUM, COUNT,
AVERAGE, etc.) over certain user-specified sample intervals. As sensor readings flow up the com-
munication tree, they are aggregated by intermediate nodes that are able to meet the requirements of
thequery(Figure.).Withoutaggregation,everynodeinthenetworkneedstotransmitnotonly
its own reading but also those of all its children. his causes a bottleneck close to the root node and
also results in unequal energy consumption, i.e., the closer a node is to the root node, the larger the
number of messages it needs to transmit which naturally results in higher energy consumption. hus
nodes closer to the root node die earlier. Losing nodes closer to the root node can have disastrous
consequences on the network due to network partitioning. Using in-network aggregation however,
every intermediate node aggregates its own reading with that of its children and eventually transmits
only one combined result. his also naturally implies that the size of the message remains constant
as it traverses from the source nodes to the root. TinyDB also illustrates how aggregation can be
 
Search WWH ::




Custom Search