2015-04-28 EECS 556 abstracts / schedule, 4-6pm 1005 EECS

Part I: Denoising

1 4:05
Image denoising using the higher order singular value decomposition
Shweta Khushu, Tanya Verma

Image denoising is a widely researched topic in the history of image
processing and continues to receive attention from researchers all over to
better the current state of the art. In this project, we implement a
patch-based method for denoising a grayscale image using the higher order
singular value decomposition (HOSVD) which achieves close to the
state-of-the-art performance. The technique stacks patches similar to a
reference patch in an image, by using a similarity index defined by a
statistically motivated criterion, into a 3D stack and performs HOSVD on the
stack. The transform coefficients are manipulated using hard thresholding and
the HOSVD transform is inverted to reconstruct the 3D stack, thereby filtering
all the individual patches. The procedure is repeated over all the pixels in a
sliding window fashion with averaging of hypotheses to produce the final
filtered image. Additionally, we discuss impact of various parameters of the
algorithm on accuracy and computational efficiency. We also augment HOSVD
denoising with a Wiener filter step, calling it HOSVD2 denoising. The results
from HOSVD and HOSVD2 image denoising algorithm are compared to other
denoising algorithms existing in the literature including NL-Means, BM3D and
LPG-PCA.


2 4:25
Texture conserving image denoising by gradient histogram preservation
Ripudaman Singh Arora, David Hong, Alfredo Bravo Iniguez

Image denoising is the task of estimating an image from a noise-corrupted
version. This is a fundamental challenge in image processing with
applications ranging from commercial photography to pre-processing for
computer vision algorithms. Though many successful denoising algorithms
have been developed, they tend to smooth fine textures and strong edges,
deteriorating the image's visual quality. To deal with this problem, some
authors have recently proposed a novel denoising method that attempts to
preserve textures by conserving the histogram of the image gradients. The
proposed method consists of two steps. The first step estimates the
gradient histogram of the original (noise-less) image from the gradient
histogram of its noisy version. The second step denoises the image while
forcing the gradient histogram of the denoised image to match the estimated
gradient histogram. Two variants are also proposed to handle images that
contain regions with distinct textures. The authors finally provide
experimental results suggesting that the proposed method performs on par
with various state-of-the-art algorithms on the basis of PSNR and SSIM,
while outperforming these methods on a qualitative comparison of the
texture. In this report, we reproduce a few of their experimental results
with a more quantitative measurement of the texture preservation and
propose extensions to handle image restoration and non-Gaussian noise.


Part II: Deblurring


3 4:45
A general framework for regularized, similarity-based image restoration
Yu Chen, Yumeng Shang

To deal with image restoration problems, A. Kheradmand and P. Milanfar
developed an iterative graph-based framework for image restoration
based on a new definition of the normalized graph Laplacian. They proposed a
new cost function, which consists of a new data fidelity term and
regularization term derived from the specific definition of the normalized
graph Laplacian. The normalizing coefficients used in the definition of the
Laplacian and associated regularization term are obtained using fast symmetry
preserving matrix balancing. The algorithm comprises of outer and inner
iterations, where in each outer iteration, the similarity weights are
recomputed using the previous estimate and the updated objective function is
minimized using inner conjugate gradient iterations. To test the general
applicability of this algorithm on image restoration problems, we do
experiments on different kinds of blurs, such as Gaussian blur, box average
blur, out-of-focus blur, and motion blur, and compute both the PSNR and SSIM
values of the deblurred images. Also, we compare some of the results with
other deblurring methods. Besides, an experiment on image sharpening is
implemented. Most of the experimental results verify the effectiveness of the
proposed algorithm on different image restoration problems.


4 5:05
Parametric blur estimation for blind restoration of natural images:
linear motion and out-of-focus
Soumyanil Banerjee, Ajay Vasudevan

The project studies a method to estimate the parameters of two types of image
blurs, namely, linear motion (which is when the camera moves in a uniform
linear motion at an angle and is characterized by a line of length L and angle
theta) and out-of-focus (which is when the camera is out of focus and is
modeled as a uniform disk characterized by its radius which specifies by how
much the camera is out of focus), for blind restoration of images. We make
assumptions such as the images are natural images, i.e., they have an
approximately isotropic power spectrum and their spectrum follows a power law
decay with spatial frequency. The project included implementation of Radon-d &
the Radon-c transforms, which are the modified Radon Transforms and error
calculation with modeled functions (fitting functions) to calculate the blur
parameters.


Part III: Segmentation / classification


5 5:25
Decoupled active contour (DAC) for boundary detection
Madan Ravi Ganesh, Adeline Hong, Leyou Zhang

This project proposes to replicate the work in the paper "Decoupled Active
Contour (DAC) for Boundary Detection". This paper seeks to split the energy
minimization scheme into two independent steps, measurement update (external)
and prior addition (internal), to improve both accuracy and speed of
convergence to a desired contour. This is accomplished by employing a
measurement update step, built on a Hidden Markov Model (HMM) and Viterbi
search algorithm, followed by a prior addition step based on measurement
uncertainty and a non-stationary prior. In our project, we will replicate the
work of the authors to boost the speed and accuracy of convergence to contours
whilst verifying its performance under challenging conditions. The challenging
conditions include presence of noise, high-curvature contours and random
initialization of snake. Proposed extensions include evaluation of the
algorithm over images with different types of noise,and modification of the
algorithm to enable adaptive algorithm parameter tuning based on image
gradient.


6 5:45
Sparse coding with a universal regularizer model for image classification
Tianyu Jiang, Sydney Williams

This project replicates work from "Universal Regularizers for Sparse Coding
and Modeling" in IEEE Transaction on Image Processing.  In the paper, a
universal coding framework is derived and implemented for creating a sparse
coding regularization term.  We employ this method for the application of
image classification in which images are sorted based on their sparse
representations.  Using one database texture images are classified by texture
type.  In a separate database scenery images are classified as object or
foreground in a basic detection problem.  Sparse coding classification with
the universal regularizer model proposed in the original paper is compared to
an l1 penalty function, a common sparsity-promoting regularizer.  We do not
make as strong of claims as the original paper about the advantages of the
universal regularizers over l1 regularization, however in some cases we find
that the universal model performs much better than l1 in terms of
classification accuracy.  We additionally explore the optimality of both
regularization methods for the application of image texture classification.
We examine the regularization parameter space of these models at an attempt to
tune these parameters for highest classification accuracy.  We also look at
the relationship between percent sparsity of the sparse representations and
classification performance.


7 6:05
Object detection on satellite map based on SIFT keypoints and graph theory
Kising Lee, Zhixun Zhao

High resolution satellite images can provide a wealth of valuable data such as
urban-suburban-rural boundaries, building location, road location, or
boundaries of bodies of water.  Manual extraction of this type of data can be
tedious and there is no guarantee of consistency.  An automated algorithm;
however, can be implemented to effectively extract the desired information.
Unfortunately, construction of such an algorithm is not straight forward if
standard techniques of image processing and pattern recognition are used
because the aforementioned objects share similar patterns that standard
techniques will have difficulty distinguishing between.  Thus, we intend to
use a scale invariant feature transform (SIFT), which has the advantage of
being invariant under various image, in conjunction with graph theoretical
tools to identify building that are present in a given image.


Presented previously:

8
Blur estimation methods for blind deconvolution
Brian Gonzalez, David Hiskens, Arvind Prasadan