Database Reference
In-Depth Information
￿
Utilizing unlabeled data to enlarge training sets, in the same spirit as active lean-
ing , for relevance feedback to increase learning capability and fast convergence.
￿
Minimizing relevance feedback iterations so that there is no requirement of
transmitting training images and video files over the distributed multimedia
database (i.e., internet, cloud, and peer-to-peer databases), reducing the required
transmission bandwidth.
The topics addressed in this chapter are as follows:
Section 3.2 presents a framework of pseudo relevance feedback and a relatively
new approach to the problem of unsupervised learning, the self-organizing tree map
(SOTM). These are the essential tools for the implementation of the automation. The
SOTM is a new member within the family of generative, self-organizing maps. Its
architecture is based on dynamic self-organization and is suitable for data clustering
in the current application.
In most of the centralized and distributed database systems, multimedia files
are stored in the compressed formats (e.g., JPEG and JPEG2000). Thus, real-
time indexing and retrieval of these files requires algorithms that can process
the compressed data without full decompression of files. It is necessary to adopt
compressed domain indexing to accomplish the low computational complexity at
run time. Section 3.3 applies the pseudo relevance feedback method to the energy
histogram features in the compressed domain of the JPEG coder, as well as other
types of compressed domain features extracted from the wavelet-based coders.
Section 3.4 explores the automatic retrieval framework by incorporating the use
of knowledge to produce some levels of the equivalent classification performance
used in human vision. The region of interest characterizes perceptually important
features, and offers a weighting scheme for the unsupervised data classification.
Finally, the automation for video retrieval will be presented in Sect. 3.5 The
spatial-temporal information of videos needs be properly captured in the indexing
stage. Then, an adaptive cosine network is applied to implement pseudo-relevance
feedback, as the network's forward-backward signal propagation, to increase
retrieval accuracy.
3.2
Pseudo Relevance Feedback Methods
3.2.1
Re-ranking Domain
Pseudo-relevance feedback is referred to as blind relevance feedback. The phi-
losophy behind this method is that the retrieval system is able to make use of
unlabeled data to improve the retrieval performance from the initial search results.
The essential task is to obtain a set of pseudo labels (i.e., the label of samples that
have been evaluated by a machine, not human users) for training relevance feedback
algorithms. Obtaining meaningful and effective sets of pseudo labels is challenging
and has been researched extensively.
Search WWH ::




Custom Search