INTRODUCTION TO CONTENT-BASED VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 10

the seemingly low-level features of projected motion from the video se-
quence that our systems use. However, projected motion often implies
high-level information such as object membership. For our algorith-
mic design, motion estimation exemplifies a video processing tool that
merges both high-level understanding and low-level computation.
Throughout this topic, we use estimates of projected motion to ro-
bustly detect and localize objects. Motion estimation plays a crucial
role in our systems because it provides a robust discriminant to find ob-
ject boundaries and the informational bootstrap step in our video object
extraction system. Instead of a simple function I ( x, y, t ), motion esti-
mation describes the video sequence as set of moving objects and forms
an alternate description of the video sequence as sets of pixels that are
correlated through time.
Our algorithmic work in our systems has much of same structure as
traditional mot ion estimation algorithms. Motion estimation algorithms
are themselves a type of content-based processing, since projected mo-
tion often can only be resolved when content is taken into account. The
computation of motion field in Section 2.7 uses many of the same tech-
niques of our video object extraction in Chapter 4 such as its energy
function formulation, its counterbalanced optimization and its iterative
updating scheme.
NEURAL NETWORKS / ADAPTIVE SIGNAL
PROCESSING
Although the two previous sections have concentrated upon deriving
visual artifacts, comprehension of the video sequence depends upon the
unseen ( a priori knowledge) as much as the seen (the video sequence,
I ( x, y, t )). Neural networks and adaptive signal processing form the core
algorithmic technology for robust processing of spatial and temporal
features, since they have the key properties of content-based processing
systems:
1. adaptability to the content within a given data set,
2. the ability to integrate high-level direction into low-level processing
by their connectivist computation,
3. and the ability to adjust to the long-term definition of content through
learning.
Within the context of our video processing tasks, neural networks and
adaptive signal processing provide a good solution to content-based anal-
ysis problems since they closely mimic human computational techniques.