Image Processing Reference
In-Depth Information
27.1 Scalable Analysis/Introduction
Historically, scientists have relied on conditional statistics, applied globally, to effec-
tively reduce large-scale data to manageable proportions. While pre-computing a sin-
gle set of structures is feasible, appropriate parameter choices are not always known
a priori, and exploring the parameter space by extracting many sets of features is
becoming infeasible due to massive data sizes. Furthermore, traditional statistics
typically provide only global averages rather than per-feature information, making
simple queries such as how many features exist overall, difficult to answer.
This chapter summarizes a new integrated analysis and visualization framework
that enables a free choice of feature parameters and conditional sub-selections;
providing the capability to interactively produce a wide range of diagnostic plots
equivalent to the processing of the entire data set. Furthermore, the statistics viewer
is cross-linked to a visualization of the corresponding three dimensional structures,
enabling selection of (sets of) features on either end. The feature visualization
employs a specialized volume rendering technique optimized for sparse, dynamic,
and binary segmented volumes.
Instead of extracting a single set of features, we compute a multi-resolution hier-
archy, capable of representing features for different parameters and at various scales.
In a single pass over the original data we pre-compute a large variety of statistics
for the finest resolution features. At run time the user selects parameters resulting
in a set of features whose properties are aggregated on-the-fly, allowing the user to
explore an entire family of feature definitions without accessing the original data. By
pre-computing statistics for a base set of features, and providing the user with sev-
eral multi-resolution hierarchies to explore, our system provides significantly greater
flexibility in the analysis process than the typical range queries of indexing schemes.
Additionally, the run-time aggregation avoids much of the cost of re-computing sta-
tistics for each set of features. As a result, our approach delivers the flexibility of
extract-and-analyze techniques while allowing for interactive exploration of large
data on a commodity desktop machine.
27.2 Augmented Feature Families
One of the basic concepts of our framework is the notion of a feature family .Given
an algorithm to define and extract features of interest corresponding to a parameter
p , a feature family is a one-parameter family that for every possible parameter p
stores the corresponding set of features. While any feature definition can be used
to create a feature family by exhaustively pre-computing all possible features for
all possible parameters, many popular algorithms naturally produce nested sets of
features for varying parameters. For example, clustering techniques progressively
merge elements [ 4 , 14 ] and a threshold-based segmentation creates increasingly
larger regions [ 3 ]. In these examples all features can be described by a collection
 
Search WWH ::




Custom Search