The primary motivation of this project to develop an organized, easily searchable database of bio-molecular images that will support complex search queries (see Bio Image Informatics Project). Pattern recognition and learning tools will enable biologists to tackle complex issues by processing vast amounts of data. First, we are developing new data models for biological data that are sensitive to its multimodal characteristics. Also, we are investigating learning tools for spatial relationships in retinal image set to pattern specific cell types within tissue and discover sub-cellular and cellular pattern of protein distribution. We are learning the protein distribution patterns in a cell tissue in order to visualize patterns and support complex querying.
We are designing an efficient interactive system that optimizes the complexity-accuracy tradeoff in multimedia similarity search. We propose a joint optimization techniques to design combined clustering and compression in the feature space as a framework to enable direct optimization of the search complexity-accuracy tradeoff. We are also deriving a data cluster approximation in the original space for kernel-based relevance feedback updating schemes, by either using only representative cluster features for approximate search based on their characteristics, or by estimating data kernels in the mapped linear space.
We propose a modified texture descriptor that has comparable performance, but with nearly half the dimensionality and less computational expense. Furthermore, it is easy to compute the new feature using the old one, without having to repeat the computationally expensive filtering step. This gain is based on a claim that the distribution of (absolute values of) filter outputs have a strong tendency to be Rayleigh. We also propose a new normalization method that improves similarity retrieval and indexing efficiency, based on filter statistics behavior along feature dimensions.
We address the problem of repetitive nearest neighbor search in relevance feedback presence. Relevance feedback learning is a popular scheme used in content based image and video retrieval to support high-level "concept" queries. Similarity or distance matrix is updated during each iteration of the relevance feedback search and a new set of nearest neighbors are computed. This repetitive nearest neighbor computation in high dimensional feature spaces (in hundreds) is expensive, particularly when the number of items in the data set is large (hundreds of thousands). In this context, we present a scheme that exploits correlations between two consecutive nearest neighbor sets and significantly reduces the overall search complexity. We show that vector quantization based indexing structures can support relevance feedback and suggest a modification to an existing nearest neighbor search algorithm to support relevance feedback for the weighted Euclidean and quadratic distance metric. Detailed experimental results are provided using image datasets.
A visual thesaurus provides the conceptual framework for observing perceptual events. Spatial relationships between these events are computed for the dataset and represented by Spatial Event Cubes (SECs). The construction and analysis of Spatial Event Cubes scale with data dimension and dataset size. The SEC representation provides appropriate input for a set of data mining applications and high-level semantic analysis. Early examination shows that the SECs are a good visualization and guidance tool for clustering in high-dimensional spaces. We also introduce perceptual association rules, a novel extension of traditional association rules, are used to distill the frequent perceptual events in large image datasets in order to discover interesting patterns. The focus is on spatial associations although the method is equally applicable to associations within or between other dimensions; i.e., spectral, or in the case of video, temporal. A primary contribution is the derivation of image equivalents for the traditional association rule components, namely the items, the itemsets, and the rules. The proposed approach is modular, consisting of three steps that can be individually adapted to a particular application. The higher-order associations and rules are determined using an adaptation of the Apriori algorithm. The proposed approach is applied to an aerial video dataset to demonstrate the kinds of knowledge perceptual association rules can help discover.
Summer Intern project at Multimedia Asset Management Group, HP Labs, Palo Alto,
Summer 2000.
Proposed, patented and partially implemented an effective multimedia classifying
system. This system comprises of different modules that enhance its efficiency,
and categorizes non-textual subject data on the basis of descriptive class
labels.
Final project for "Digital Image Processing" course, Spring 2000, Prof. B.S.Manjunath,
Implementing and modifying of existing image segmentation algorithms.
Final project for "Neural Networks" course, UCSB, Winter 2000, Prof. B.S.Manjunath,
Explored possible use of Radial Basis Functions to improve Image Database search engine.
Final project for "Indexing in Multimedia Databases" course, UCSB, Fall 1999, Prof. Ambuj Singh
Implemented existing dimensionality reduction algorithms and applied them to large texture datasets.
Evaluated performance with respect to scalability, cost, and error rate of the output.
Final project for “Multirate Signal Processing” course, UCSB,Spring 1999; Prof. Chuck Creuseire
Presented basic principles of cyclic multirate systems and filter banks. Introduced additional
freedom that cyclic system offer with respect to traditional filter banks.
Final project for “Advanced Digital Signal Processing” course, UCSB,
Winter 1999, Prof. Sanjit Mitra
Mapped long running convolutions into smaller ones by using filter banks based on an
aperiodic convolution algorithm. Good tradeoffs among computational complexity,
input-output delay and system architecture are achieved.
IAESTE internship, University of Campinas, Brazil, November 1997 - February 1998.
Development and application of an algorithm for effective storage and visualization of the data during the
Holter ECG monitoring using MPEG2 video compression under guidance of Professor Dalton Arantes.
Calthech Undergraduate Research Fellow, Jet Propulsion Laboratory, Pasadena, CA,
Summer 1997.
Documented and simulated the history and development of Probe Entry Site(PES)
Hot Spot motion on Jupiter over the two years period( March 1995 - May 1997).
Simulation was presented on The 29th Annual Meeting of the American Astronomical
Society, MIT, Cambridge, MA, 1997. Proved that reconstructed image, interpolated
between higher spatial resolution images, is a reasonable approximation to the
real, blurred image. Improved the flexibility of the main user interface program
for data reduction and created a functional form for data preprocessing for
alternative environments. Created a procedure that transfers image cylindrical
map to a disk map, in order to increase signal to noise ratio.
Diploma thesis project, University of Belgrade, Spring 1998.
Investigated and simulated the application of adaptive filters to noise reduction in CDMA
receivers under supervision of Professor Zoran
Dobrosavljević.
Joint research project with Medical Institute, University of Belgrade, Spring 1998.
Segmentation and classification of the abnormal cells obtained from digitalized tissue images using
a combined segmentation methods: edge detection, two-level threshold and
region growing techniques