Implementation of MPEG-7
visual descriptors in Java
Monika Kostolná
Objective
 Reimplementation of chosen MPEG-7
descriptors in Java
◦ Scalable Color Descriptor
◦ Edge Histogram Descriptor
 Application of MESSIF framework (DISA)
 Using potential of java libraries
 Comparison with C++ Extractors
Scalable Color Descriptor
 Represents the color element in the image
◦ HSV scheme
 Creation of the uniform histogram
 Haar transformation
◦ Sums ~ low pass filters
◦ Differences ~ high pass filters
◦ Compression with no data loss
 Scalability in the context of
◦ Number of output coefficients
◦ Number of bit planes discarded
Example
 Inputs or outputs can be different
◦ The uniform histogram creation
◦ The length and the bit depth of the output
are defined by coefficients
 Example:
 NumberOfCoefficients = 8
 NumberOfBitPlanesDiscarded = 3
Edge Histogram Descriptor
 Represents the texture and the shape
elements in image
 Input divided into 4x4 sub-images
◦ For each sub-image is calculated an
histogram with 5 types of edges
 Vertical
 45-degree diagonal
 Horizontal
 135-degree diagonal
 Non-directional
 Final histogram of 80 bins (4x4x5)
Example
 2 images with different locations of edges
1 10 20 30 40 50 60 70 80
1 10 20 30 40 50 60 70 80
Advantages of Java implementation
 Usage of Metric Similarity Search
Implementation Framework (MESSIF)
◦ Simplification of the work in metric space
◦ Direct connection to application using MESSIF
◦ Used for
 Comparing C++ and Java implementation on
andromeda
 Extraction output
 Testing
Advantages of Java implementation
 ImageJ API
◦ Open source program
◦ ImagePlus, ImageProcessor
 Contains pixel data of 2D picture
◦ ImageConverter
 convertToGray8
 Maven
◦ Tool for
 Dependency management
 Project building
◦ Dynamically downloads required java libraries
Comparison with C++ extractor
 Portability
◦ C++
 Platform dependent build required
◦ Java
 Platform independent (portable)
 Creation of native code executable on any machine
 Code readability
◦ Original C++ source code
 Very long and difficult to maintain
◦ Created Java source code
 Simple and documented methods
 Conventions following
Comparison with C++ extractor
 ScalableColor
◦ 50 small images
 C++: 613 ms
 Java: 775 ms
◦ 50 big images
 C++: 3420 ms
 Java: 5506 ms
 Future optimization
possibilities
 EdgeHistogram
◦ 50 small images
 C++: 203 ms
 Java: 280 ms
◦ 50 big images
 C++: 822 ms
 Java: 1248 ms
Thank you for your attention.