|
|
- Info
PDF-Files
-
diploma thesis
-
wscg05: posterpaper
-
wscg05: poster
-
cvww05
-
ppgt05
-
CVWW06: Paper
-
bmvc07
-
bmvc07: poster
-
tamee
-
sensors and actuators
-
micro-colony
-
Incremental, Robust, and Efficient Linear Discriminant Analysis Learning
-
This thesis is focused on Linear Discriminant Analysis (LDA), which is a
subspace learning method. LDA is employed for appearance-based object
classification. The standard LDA needs all training data to be given in
advance in order to construct the subspace. This type of learning is termed
batch learning. But in general, not all data is available at the same time.
In order to avoid storing the complete data it is necessary to process
learning samples as soon as they become available and discard them
immediately afterwards. Consequently, instead of a new subspace construction
an approach is desirable directly adapting the current subspace to represent
the old as well as the new data. We call this type of updating incremental
learning. Furthermore, consider the task to generalize the classification
from individual classes to joint categories. Since the original data is no
longer available for this retraining only the data representations can be
used. We develop solutions for these tasks actually trying to simulate the
humans ability to adapt to a changing world. Another power of the human
vision system is that it can easily compensate for missing information. But
if there is wrong information described in the subspace classification will
fail. Consequently, a reliable noise detection is needed, and the ability to
construct a subspace solely from partial images is required as well. An
answer to this robustness problem is given for both issues. That is, batch
and incremental learning are adapted to handle non-Gaussian noise, which
occurs in particular due to occlusions and missing pixels. Finally, it seems
that humans have no restriction of the number of learnable classes. The
performance of LDA on the other hand decreases with a growing number of
classes. Besides, classes with a large variability in appearance cannot be
handled properly. Therefore, an adaption of LDA such that it can deal with
these restrictions is worthwhile and presented in this thesis. In order to
develop solutions for the mentioned LDA learning problems it proved that a
combination of reconstructive and discriminative information provides the
necessary basis. Exploiting the properties of both types of information we
present algorithms for incremental updating and robust learning for both
training scenarios (batch and incremental learning). In addition, we provide
an alternative to the single LDA subspace approach such that even a very
large number of classes can be classified satisfactorily. All claims are
evaluated exhaustively on different datasets of different size and level of
difficulty.
|
