General notes
Before going into the details of the possible projects it is worth
pointing out some general tasks that will be required no matter what
the specific project is.
- There is no single wavelet transform and a major part of the
project will be to determine what transform is most appropriate
for the specific application at hand. This choice should be
clearly motivated in the project report and justified based on
the characteristics of the problem.
- To select a wavelet transform it will be necessary to
determine:
- Filters to be used
- Number of levels of decomposition
- Type of decomposition (dyadic/wavelet packet)
- Are the same filters used at every level?
- Is the transformation changed spatially (i.e. different signal segments are transformed differently)
- Is the same transform used for each signal, or is a signal adaptive transform used?
Music/Speech Discrimination
Separation of music from speech
is useful in automatic speech recognition applications. It can also be
used to identify between songs and advertisements in radio
broadcasts. It can also be useful in audio databases where it is
desirable to have content-based indices that will allow a user to
query specific types of data.
The characteristics of music and speech are different as
they have different sources, hence by the appropriate choice of
features we should be able to distinguish between music and speech.
The references list a variety of features that can be used to identify
the music and speech segments. Emphasis in this project should be on
the use of wavelet features to do the above.
Segments of speech
containing both music and speech will be provided, the goal is to
segment the audio into separate portions containing music and speech.
References
1
2
Contour Representation
In applications like computer graphics, image analysis, it is required to
describe curves to facilitate further processing. Approximations of curves can
be obtained by using multiresolution analysis. The features used to represent
the curve should be insensitive to rotation, translation and scaling. Features
extracted by wavelet analysis satisfy these constraints and are suitable for
contour representation.
In this project several contours will be provided and the goal will be use
wavelets to efficiently represent them.
References
1
Texture Classification
Texture Classification involves classifying an unknown texture into one of
several classes based on known textures. The classification is motivated by
similarity between the unknown and known texture. While this classification
is relatively easy to perform when done visually, it is heavily dependent
on the features used for classification when done by a computer.
In this project the objective is to classify the unknown texture(s). Wavelet
based features should be extracted from the unknown texture and a suitable
distance metric should be used to compare it with the features of the known
texture.
Several textures will be provided as reference textures. Natural and
synthetic textures will need to classified into one of the class of the
reference texture.
References
1
De-Noising
Noise in images can be added due to camera, scanners etc. To improve the quality
of the image it is desirable to remove/suppress the effect of noise. In this
project images corrupted with additive noise will be given, the objective will
be to use wavelets to suppress the noise in the images. In order to do denoising wavelet coefficients are extracted from the image and then noise suppression is
done by hard/soft thresholding. The basic idea is that natural signals not
corrupted by noise tend to be "smooth", whereas when noise is
introduced the degree of smoothness is reduced (e.g., more coefficients
appear at high frequencies than in a noise free image.)
In order to obtain better performance, wavelet
packets and adaptive techniques should be used.
References
1
2
Texture Generation
The texture generation problem can be seen as a converse of the texture
classification problem. While some textures have some well defined
structure, others are more noise like. The basic idea here is to provide
a mechanism such that textures can be generated by generating random
values for wavelet coefficients. By choosing the appropriate random
representation and the different random models for each subband,
different types of textures can be generated.
References
Graphics
Wavelets have been very effective in representation and analysis of data in
Euclidean spaces. However in applications like
computer graphics, mapping earth
topology, we need to deal with data on complex geometries(ex surface of a
sphere).
Traditional
wavelets cannot be used for analysis of data on these complex
geometries.
Spherical wavelets(1) have been
proposed which can be used for representation and compression of functions on
spheres.
In this project graphical data on a sphere will be given and spherical wavelets
(or any other suitable wavelet) should be used for analysis of data defined
on the sphere.
References
1
2
Segmentation
For better compression, image analysis,
classification; segmentation of the image is important. Here the image
is separated into different regions such that each region is
relatively homogeneous. Wavelet based features provide a good
representation of the image and can be used for image segmentation
purposes.
Several images(natural and synthetic) with varying textures will be
given and the goal will be to efficiently segment the image.
References
1
2