Projects and Seminars


Course Projects at UCSC

Topic : K-SVD Algorithm for Denoising of Gray-Scale Images
Course : Image Processing (Spring, 2007)
Instructor : Prof. Peyman Milanfar

Sparse representation of images has been a recent area of growing interest. It finds applications in many problems in image processing. In this report we study a particular method of achieving sparse representation using the recently proposed K-SVD algorithm by Aharon et al. and how this sparse representation framework has been extended to perform denoising, as illustrated by the authors. Suitable illustrations along with the theory is detailed in this report to help in understanding of the intricacies of this method.

Report

=~=~=

M.Tech Dissertation

Topic : Image Super-resolution & Inpainting from a Single Observation
Advisor : Prof. Subhasis Chaudhuri
Examiners : Prof. S. N. Merchant, Prof. Sharat Chandran

Abstract : In this work, we deal with the concepts of image super-resolution and inpainting from only a single input image. A study of the Papoulis-Gerchberg method is presented and its applications to these domains of computer vision are demonstrated. We address the issue of its inability to deal with blurred low resolution images and show results that prove the effectiveness of our method when the low resolution image formation model is known. Another learning based approach is presented where we try to do away with such an assumption by trying to learn a filter that bridges the gap between the reconstructed and the desired high resolution image. The PG method can also be extended to the domain of image inpainting where filling-in is to be done for thin scratches. We propose alterations which lead to better inpainting in specific cases. A faster method for inpainting within this framework is also proposed by making use of different orthogonal basis functions.

We then explore the use of regularization based techniques for image super-resolution. The total variation is used as a regularizer to take advantage of it edge preserving nature. However, a direct application results in loss of textural details in the super-resolved image. We propose the use of additional data fidelity constraints to perform texture preserving super-resolution when the degradation model is available. In order to make the process robust in the presence of noise, this process is further extended by using spatially varying weighing terms in the objective function. We demonstrate encouraging results obtained using our methods that strengthen our theory and pave a way for future research.
Masters Dissertation

=~=~=

M.Tech Seminar

Topic : Control Applications in Communication Networks
Advisor : Dr. Harish Pillai
Examiner : Prof. V.R. Sule

Abstract : In this seminar I did a control-theoretic study of Congestion control in a computer network. The report is a summary of just a few of the popular controllers that are used for congestion control and the underlying theory behind their use.

Seminar Report  Presentation slides

=~=~=

Course Projects

  • Topic : Point based rendering using Surface Splatting
    Course : Advanced Computer Graphics (Spring, 2006)
    Instructor : Prof. S. Chandran
    Collaborators : Anish Chandak, Rhushabh Goradia, K Anil.

    Abstract : The aim of this project can be divided into two parts i) Formulation of point models from models in MGF file formats and ii) Rendering the scene using the point based model obtained from the first part. My work was mainly concerned with the latter part. A comprehensive summary of the work done can be found here. This work was a group project done with 3 other friends.

    =~=~=

  • Topic : DSP Application in Signal Processing of Modern RADAR
    Course : DSP and its Applications (Autumn Sem, 2003)
    Instructor : Prof. V.M. Gadre
    Collaborators : Jay Makhija

    Abstract : Advances in the field of DSP have led to considerable improvement in RADAR technology. It makes the system flexible in allowing vast number of functions to be performed with finite number of components, yet being economic. In RADAR, DSP is mainly applied in signal processing and in Data Processor Systems (DPS). In this paper we explicitly deal with application of matched filter in signal processing. Basic theory of matched filter is being dealt with, and implementation of matched filters employing pipelined FFT for the realization of high speed convolution is explained.

    Report

    =~=~=

    Detailed list will be put up later....

B.Tech Project

Topic : Braille to Text Conversion
Advisor : Dr. Ajay Ghosh (Univ. of Calcutta)
Examiner : Dr. S. Barat

Abstract : This project dealt with the possibility of converting a braille document to text form. The braille documents are first scanned and then processes by our algorithm to find out the characters in the document. Before the characters are located it is first processed to determine the amount of zooming done in the scanning process. Some simple restoration is also done so as to enhance the capability of detection and recognition. The characters are then processed to find the code and the alphabet corresponding to the read code is displayed. We show an algorithm that works quite fast and is able to convert braille documents to text with satisfactory accuracy.

Project Report

=~=~=

Other Projects

Topic : An Object Oriented Approach to Data Routing in a Network
Advisor : Dr. Suman Chakraborty (Dept. of Mech. Engg., IIT-Kharagpur)

Abstract : In this paper, a suitable generalized object-oriented formulation is devised, which can be used to find the shortest paths between machines for all types of networks. The use of object-oriented programming paradigm allows the programmer to add greater flexibility to the programs by modeling them as entities (called Objects) that have certain common properties (which are defined by the Class that they belong to). Apart from the flexibility offered to the programmer, a major advantage in Object Oriented protocols is in the teaching field. Non-object Oriented implementations make use of certain data structures which represent the parameters of a network (and thus its state at any point of time). But in Object Oriented implementation the Classes are modeled taking into account the physical parameters of a network. So it is not difficult to grasp the data structures used. Moreover, due to the use of Object Oriented methodologies, it is easier to visualize the operations in the algorithm. The same algorithm can be run without any modification, even if network characteristics and topology change. Addition or subtraction of routers and links can be achieved by simply addition and subtraction of Nodes and Edges. Additionally, for this algorithm the weighting function just needs to be changed to incorporate the requisite changes for different types of subnets. This makes it easier for the student to assimilate the concepts and metrics involved in different kinds of networks. Finally, the algorithm is aptly illustrated by means of two representative case studies.

P. Chatterjee, S. Chakraborty, An Object-Oriented Approach to Data Routing in a Network, International Journal of Industrial and Systems Engineering, to appear.

=~=~=