Feature transformation using evolutionary computation for multimedia information content characterization

利用進化計算進行多媒體信息內容特性化的特徵變換

Student thesis: Master's Thesis

View graph of relations

Author(s)

  • Chun Ip CHIU

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date4 Oct 2004

Abstract

Due to the rapid development of multimedia technology, huge amount of multimedia information has been generated in today multimedia applications. Examples include the advancement in digital photography technology, 3-D graphics and virtual reality technologies. This, in turn, effective as well as efficient management of these systems has become a critical issue. It has been well documented that simple textual annotations are often ambiguous and inadequate for multimedia, information retrieval or classification. Consequently, an intensive research area has emerged called content-based retrieval (CBR) in which the main idea is to retrieve information semantically. In other words, the retrieval or classification is done based on the underlying content of the multimedia entity. Feature extraction and indexing are the fundamental elements for CBR applications. Generally, features are extracted to represent the underlying content and use as signatures for retrieval or classification purposes. For example, features such as color, shape and texture are usually used to capture the characteristic of image content for contented-based image retrieval. Normal vector orientation angles of the individual polygons in the 3-D model are also usually used as features for characterizing 3-D models. Besides, form of representation of extracted features is also an important issue. Histogram has been widely used for such propose. One of the main reasons is that it can be computed easily and efficiently. In addition, histogram is able to summarize the statistics of the extracted features that can use to reduce data dimension. Typically, it is used as feature vector for retrieval or classification purposes. However, approximate representation of some extracted features which in the form of histogram may not be able to characterize the underlying content of the multimedia information accurately. Therefore, we expected there should be a certain transformation that able to transform the histogram in such a way that the content characterization capability can be improved. With such expectation and a large search space but without prior knowledge about the form of transformation, we investigate the effectiveness of applying Evolutionary Computation to search for such suitable transformation. Specifically, we chose Genetic Algorithms (GA) and Evolution Strategies (ES) for the search process. To evaluate the approach, we focus on multimedia information categorization problem. We apply the approach to compress-domain images and 3D head models categorization problems.

    Research areas

  • Multimedia systems, Evolutionary computation