The University of Tennessee
Knoxville, TN 37996-2100
IRIS Research Staff |
Vivek Agarwal |
|
|
|
Office: |
401 Science and Engineering Building. The University of Tennessee Knoxville, TN 37996-2100 |
| Telephone: | (865) 974-9737 | ||
| Fax: | (865) 974-5459 | ||
| E-mail: | vivek@utk.edu | ||
| Personal Web Page: |
Recently number of machine learning techniques had been proposed to solve color constancy problem in computer vision. Specifically neural network and support vector regression (SVR). Neither Neural networks nor SVR had been compared to simpler statistical approaches in those studies. In my research, linear regression technique - ridge regression is presented and show that ridge regression performs better than neural network and SVR in estimating illumination chromaticity. Uncertainty analysis is performed for neural network, SVR, and ridge regression using bootstrapping. From the analysis, it is observed that ridge regression and SVR are more consistent than neural network but ridge regression has upper hand compared to SVR because it is computationally simple and fast. MS Thesis (PDF 1.56 MB) Presentation (PPT 7.0 MB)
Other Research Projects:
 |
Edge Preserving Image Restoration using L1 Norm Image restoration is an ill-posed problem in image processing. Different theories and algorithms are present to achieve image restoration from a blurry and noisy image. A area of focus in this work was to obtain high quality fast image restoration by preserve the high frequency information in the restored images using Ln norm regularization approaches, i.e., edges. Classical regularization (L2 norm regularization) smoothes out the edge information in the restored images. Total Variation (L1 norm regularization) preserved the edge information in the restored images but computation time was very high. LASSO (L1 constraint statistical approach) is extended to image restoration. The computation time of LASSO is very less, almost twice as fast as Total Variation regularization. Presentation (PPT 5.0 MB) Technical Report (PDF 0.2MB) |
 |
Discrete Nonlinear Programming for the Optimal Selection of LED Arrays In collaboration with Siemens Energy and Automation, Johnson City, Knoxville, formulated an algorithm for the design of LED traffic signals. The designed required optimal selection of sets of LEDs from a random distributed LED bins varying in three aspects, i.e., from a 260 combination of LED bins in each color. The selected LEDs are placed on circuit automatically as per design criteria. Poster (PPT 1.7MB) Technical Report and Presentation is not publicly available. |
 |
Multiscale Retinex for Color Image Enhancement Implemented the retinex theory on human vision for machine vision. The algorithms has a unique feature of enhancing the features that are not clear under different illumination conditions. The color restoration helps to restore the true color in the image with small loss of color constancy. Presentation (PPT 7.8MB) Technical Report (PDF 3.62MB) |
|
Publications: [1] Vivek Agarwal, Andrei V. Gribok, and Mongi A. Abidi, "Machine Learning Approach to Color Constancy," submitted to IEEE Transactions on Neural Networks - Letter. [Abstract] [2] Vivek Agarwal, Andrei V. Gribok, Andreas Koschan, Besma R. Abidi, and Mongi A. Abidi, "Ridge Regression for Color Constancy: A Simple Approach," under preparation for submission to IEEE Journal. [3] Vivek Agarwal, Andrei V. Gribok, and Mongi A. Abidi, "Image Restoration using L1 Penalty Function," under preparation for submission to Journal of Inverse Problems |
|