Biomedical Engineering Reference
In-Depth Information
13.2.1 Image Analysis
Basic image data contains much information for interpretation. This information
is available from the actual sensed values (i.e., the pixels). Modern sensors can
scan over time, space, and spectrum. Imaging techniques such as Atomic Force
Microscopy (AFM) provide new view points on the real world. Image processing
allows these images to be interpreted generally by humans, but also, by machines.
A fundamental action for image processing is object recognition, and the pro-
cess begins with segmentation of the image. Bisque integrates several general and
biologically specific segmentation methods which store derived information as ob-
ject shapes or statistics. Many analysis types operate over multidimensional images
looking for changes over time or space. For example, Bisque offers microtubule
body tracking and is able to measure changes to microtubule lengths. This de-
rived information can be summarized as microtubule growth analysis revealing
fundamental properties.
13.2.2 Indexing Large Image Collections
Any image repository, be it a photo sharing site, a biomedical image database, an
aerial photo archive, or a surveillance system, must support retrieval, analysis, com-
parison, and mining of images. Over the years, image datasets have gained promi-
nence in scientific explorations due to their ability to reveal spatial information and
relationships not immediately available from other data sources. Decreasing cost
of infrastructure and the emergence of new tools has brought multimedia-based
social networking sites and image-based online shopping sites into prominence.
Existing query tools are being used in aerial photography and face recognition
for automatic image annotations [24, 30]; in astronomical satellite images to ac-
curately detect point sources and their intensities [12]; and in biology for mining
interesting patterns of cell and tissue behavior [5, 9]. The development of function-
ally enhanced, efficient, and scalable techniques for image retrieval and analysis
has the potential to accelerate research in these domains and open new frontiers
for commercial and scientific endeavors.
Two problems have plagued the utility of content-based querying of multime-
dia databases: (1) the definition of similarity between images, and (2) the restriction
of distance measurement over semantically meaningful pairs of objects and subim-
ages. With regard to the former, a number of distance measures have been proposed
such as Mahalanobis distance [25], Earth Mover's Distance [30], and even learning
the appropriate distance measure [31]. The eventual goal is to identify a signifi-
cant match for a pattern of interest in a query image. With regard to the latter
problem, segmentation has been used to define the meaning and context of objects
and segmentation-based image retrieval has been proposed. However, questions
about the scalability of these techniques remain. We need techniques that can find
a meaningful subregion without much user input or extensive image analysis.
Querying based on EMD:
Existing distance measures for querying and mining
complex images, such as the Lp norms, are insufficient as they do not consider
the spatial relationship among the objects. For biological images, the spatial loca-
tion is important for whether two images should be considered similar. The Earth
Mover's Distance (EMD) [30] captures the spatial aspect of the different features
Search WWH ::
Custom Search