This demonstration applies the current MARS back end prototype
to the field of remotely sensed data. The MARS prototype is better
described elsewhere, but essentially consists
of a search engine capable of searching images based on their content.
To this end a series of image processing tools are used to extract desirable
features from images so as to query based on them.
MARS provides a simple querying interface taken form the
Information Retrieval community. Currently MARS supports the boolean retrieval
model in which queries over images can be combined via the boolean operators
AND, OR and NOT. Each simple term is formed by specifying
the desired measure (i.e. color similarity, texture similarity, etc.) and
an identifying number of the query image. MARS then computes for each
simple term all matching images and the degree of the match and then combines
them according to one of two criteria: distance and probability.
Project description
- Database Size
For this demonstration, we have 30 meter satellite imagery of the Fort
Irwin Area. Our data is composed of seven bands described here:
- Band 1 - (0.45-0.52 micrometers) Water penetration
- Band 2 - (0.52-0.60 micrometers) Visible green
- Band 3 - (0.63-0.69 micrometers) Chlorophyll absorption
(vegetation detection)
- Band 4 - (0.76-0.90 micrometers) Soil-crop, land-water
contrast
- Band 5 - (1.55-1.75 micrometers) Crop and soil moisture
- Band 6 - (2.08-2.35 micrometers) Discrimination of rock
formations
- Band 7 - (10.4-12.5 micrometers) Thermal infrared
This data covers 50 x 50 kilometers and we chose to divide
it up into one by one kilometer regions. These image subsets are then used
to extract the required features and stored in a database. Additionally,
we have elevation data for the same area. This was used to
construct the colored image shown upon entry and was height color coded
according to a fairly standard color scheme for elevation data.
- Modules Of This Project
This project can be divided into 4 main components
- Back End Algorithm (Image Feature Extraction)
The algorithms that extract images features.
The other algorithm in use is a combination of the first three moments
of the intensity level of each pixel. Other algorithms are available in
the back end query engine and feature extractors, but were not used for
this demonstration.
- Back End Query Engine
The feature extraction described above is done off-line and once to build
the database. The query engine then uses these databases to process queries
submitted from the user interface.
- Front End Interface
We have a GUI and the demo is accessible from the web. Although this interface
covers the back end query engine, it serves as a conduit for the user to
access it.
The user interface constructs a query which is then submitted to the query
engine. The user can see the query at different stages of completion and
just before submission.
- Relational Database Design
The image features extracted are then stored in an image database
- User Interface Output
The User interface outputs a color coded sequence of images. The first
one shown is the above mentioned color coded by height data. From then
on, a series of images representing individual bands are shown. Each of
these images corresponds to a band that was selected by the user.
Each of these images has a colored overlay ranging from red to blue. Areas
colored are good matches, the red being the best and the blue the worst,
but still better that any non colored area. The user is given the option
to zoom in on one of the images to see more detail. A sample output is shown
here, click it for a larger view.
|
