W12 EECS 556 final project abstracts for 2012-04-16
and approximate presentation time schdule (15 min talk + 5 min q/a)


01 3:00-3:20
Nonrigid medical image registration
Kibum Bae, Jean Young Kwon, Mooyoung Shin, Yiying Zhu		(BEFORE 5PM)
: 
Image registration is a fundamental task in image processing techniques
used to match two or more image with respect to corresponding geometric
locations. Medical image registration plays an important role in many
clinical applications, in that it could improve the way of comparing and
combining multiple images acquired from subjects at different times or
different modalities.
In this project, we implemented a multimodality medical image registration
in the chest using nonrigid B-spline deformations drawn on the work of
Mattes and Haynor [1]. We built a system which fuses images from a modality
with high anatomic detail (CT) with images from a modality delineating
biological function with low anatomic detail (PET). This process involves
the concepts of interpolation, transformation, deformation, similarity
metric and optimization. We analyzed the registration process by varying
the input parameters and similarity metrics. For the purpose of
quantitative analysis, we compared the alignment of lung boundaries. 
[1] D. Mattes, D. R. Haynor, H. Vesselle, T. K. Lewellen, and W. Eubank,
"PET-CT image registration in the chest using free-form deformations," IEEE
Trans. Med. Imag., vol. 22, pp. 120-128, Jan. 2003.


02 3:20-3:40
Image mosaicing
Ridvan Eksi, Theresa Mahoney, Wan Huang, Yue Hou
:
Abstract:
In image mosaicing, multiple small photos demonstrating limited views can
be stitched together to form a panoramic image. In our project, we
implement an automatic image mosaicing system. Photos taken from a panning
camera are our input objects and a stitched image with a wider view is the
output. Two input images are stitched at a time, one of which is the static
or reference image and the other is the sensed image waiting to be
transformed. The images go through pre-processing, where features are
detected and correspondence points are selected so that the images are
registered. Using these point pairs, a transformation model is established
and the sensed image is transformed. In our project, we chose the
Projective model so that various transformation types (scaling, rotation,
distortion, etc.) are taken into account. Some non-rigid models are also
investigated to provide a comparison. The two images are then arranged onto
canvases with the same coordinates so that image fusion, using grey level
averaging and spatial frequency methods, can be made accordingly. The
images are fused and compared with the original reference image using the
minimum RMS (Root Mean Square) Error quality metric for comparison of the
various methods tested. Alternatively, in the image registration and
transformation part, another method - the Fourier-Mellin transform based
method is used to compare and contrast with the traditional feature-based
method, which is suitable for images with rotation, displacement and
scaling with respect to the reference image. However, this method is
limited in that only sensed images with small rotation and scaling factor,
and considerable overlapping with the reference image can be recovered.
Quantitative analysis will be performed to report when this method is
effective.


03 3:40-4:00
Image super-resolution via sparse representation [yang:10:isr]
Yu Cheng, Ryan Garrone, Steven Joseph, Haixin Li, Renzhi Qian	(BEFORE 5PM)
:
In many imaging applications, resolution is limited by the dimension and
number of pixels used to capture the light. There are, however, clever ways to
achieve high resolution images by capturing only low resolution images. These
methods are called "super-resolution" techniques, which primarily consist of
multiple and single image super-resolution. This project investigates the
single image case. The algorithm operates on overlapping blocks of pixels and
uses a carefully trained, over-complete, dictionary of high and low resolution
patch pairs learned from natural image patch samples. Specific image classes
with more structured data, such as faces, allow for even better resolution
performance. Both the general and specific image class cases are addressed in
this work, and broad range of comparisons are carried out in Matlab.


04 4:00-4:20 (KLA-Tencor 1st place project prize)
Robust super resolution using the method of multipliers
Alan Chu, Mai Le, Matthew Muckley, Feng Zhao
:
Super resolution methods using L1 data fit terms and total variation
regularizers have been shown to be more robust than methods using L2
data fit terms in the presence of non-Gaussian noise models, such as
salt and pepper noise. However, few algorithms exist that are designed
for L1 data fit terms. In addition, the super resolution problem can
be complicated due to the combined presence of downsampling, blurring,
and motion. We designed an algorithm based on the augmented Lagrangian
method to solve the original constrained problem by introducing a
Lagrangian cost function and auxiliary variables. In a number of
cases, iteratively minimizing the Lagrangian cost function is both
simpler and faster than solving the original constrained problem. By
using prior knowledge to formulate exact update steps and by
interleaving variable updates with Lagrange multiplier updates, we
were able to optimize the algorithm's performance in a super
resolution context. Our results demonstrate increased robustness over
previous methods in the context of ill-posed super resolution problems.


** 4:20-4:30 break


05 4:30-4:50
Image denoising: A comparison of multiple noise suppression methods
Irfan A. Dar, Abbas Sohrabpour, Hedieh A. Tamaddoni, Angelo W. Wang
:
A lot of effort in any signal processing system is put into suppressing noise.
The noise is usually modeled as additive white Gaussian with zero mean.
Historically, efforts such as finding optimum (in MSE sense) linear shift
invariant filters has led to the introduction of Wiener filters as opposed to
simple low pass filtering (in the Fourier domain). Many other methods such as
adaptive filters and nonlinear methods such as median filters have also been
proposed. The introduction of Wavelet denoising methods by Donoho has sparked
a set of new investigations into the matter. In this project, three categories
of denoising methods being the Non Local Means, Anisotropic Diffusion
Filtering and Wavelet based denoising methods, have been implemented and
compared. In addition to the traditional MSE evaluation of the different
methods, we have fed in the denoised images into a segmentation system to
compare the segmentation accuracy of different denoising methods. We have
focused on denoising MRI images of the brain. The segmentation code will
segment the white and gray matter of the brain. Monitoring the gray/white
matter degeneration helps physicians to assess the development of diseases
such as Multiple Sclerosis (MS) and thus take appropriate measures. As a
result, being able to reduce noise and thus increase segmentation accuracy can
be a potential application of denoising in medical images.

06 4:50-5:10
Computer aided diagnosis for lung nodule using image segmentation
Ke Hu, Zhihao Liu, Shili Xu, Xi Zhang
:
Segmentation methods has been widely used in Imaging-based Computer Aided
Diagnosis.  In this project, we propose to apply multi-stage image
segmentation algorithms to segment the lung nodules out from the dataset given
by the Lung Image Data base Consortium (LIDC). The whole project mainly
involves two-stage segmentations.  The first stage does lung segmentation.  We
propose to use methods based on thresholding and boundary tracing.  The second
stage segmentation is about automated nodule segmentation, on top of the
refined lung region from the first stage.  We propose to use a classification-
based image segmentation method.  The performance of each stage is evaluated
and the proposed segmentation algorithms achieve decent performance.


07 5:10-5:30 (KLA-Tencor 2nd place project prize)
KEG tracker: Local tracking leveraged by global constraints
Chen Feng, Jiwei Cao, Juchuan Ma
:
Abstract:
In this project, we propose a new registration algorithm and computing
framework, the KEG tracker, for estimating a camera's position and orientation
for a general class of Augmented Reality (AR) applications.  By studying two
classic natural marker based registration algorithms, Homography-from-
detection and Homography-from-tracking, and overcoming their specific
limitations of jitter and drift, our method applies two global constraints
(geometric and appearance) to prevent tracking errors from propagating between
consecutive frames.  The proposed method is able to achieve an increase in
both stability and accuracy, while being fast enough for real-time
applications.  Experiments on both synthesized and real-world test cases
demonstrate that our method is superior to existing state-of-the-art
registration algorithms.  Also, in the second stage of this project, we
explored two machine learning algorithms, FERNs and P-N learning, as methods
of fast re-initialization of our tracker.


08 5:30-5:50
Video completion with object track implementation under two scenarios
Sheng-Yang Hung, Cheng-Che Tsai, Zhongying Liu, Hua Feng
:
Image completion is the problem of automatically filling in part of missing
images or replacing unwanted pixels by using information from surrounding
area. Video completion can be considered as an extension of 2D image
completion by viewing as a serious of 2D images. However, video completion
is more challenging than image completion. Simply completing video
sequences frame by frame using image completion methods leads to an
inappropriate result. Since the human eyes are sensitive to object's
motion, any temporal inconsistency of structure or texture will be percept
by human eyes. Thus current existed image completion needs modification to
apply on video completion. To simplify the system, we will currently focus
on the still camera or surveillance camera. In our implementation, we focus
on two scenarios, first is get rid of any obstacle when the object is
attempting to pass behind the m. The other is remove the moving object
without ruining the background. We further compare the performance of
different image completion approach under these two scenarios.


** 5:50-6:00 break


09 6:00-6:20
iMouse: Iris control of mouse via active appearance model and gaze estimation
Gudur, Madhu Sudhan Reddy; Guo, Shijia; Pan, Ren; Wu, Ziqi
:
In human-computer interactive systems, iris movement plays an important role
in expressing the desired actions of the user. In this project, the goal is to
control the cursor actions according to the movement of the eye. Active
Appearance Model (AAM), a statistical model of the shape and grey level
appearance of the object computed via Iterative Procrustes Algorithm and
Principal Component Analysis, is used to extract the corresponding properties
of the eye corners, iris center and nose from the training images. These
properties are used in detecting the contours of eyelids and the iris. A
geometrical model is then used to estimate the gaze direction of the user,
which is then translated to identify the point-of-regard (viewing location) on
the screen. The accuracy of AAM in detecting the facial features is assessed
using Receiver Operating Characteristic (ROC) curves and the error between the
estimated and true point-of-regard is evaluated.


10 6:20-6:40
An optimal combination of watermarking and halftoning for grayscale images
Xiang Cui, Gen Lv, Baiyu Wang, Qirui Wang
:
This project attempts to find an optimal combination of watermarking and
halftoning for grayscale images. The process has three steps. First using a
continuous-tone image f[n,m] to get a watermarked version g[n,m]. Then the
watermarked image is converted to a halftoned image f~[n,m]. Lastly, the
halftone is printed out and scanned back by a flatbed scanner, denoted as
g~[n,m]. In watermarking, we use block-based spread spectrum watermarking
since this method has been proved very robust to jamming and interference
compared with other schemes in the domain of watermarking. In halftoning,
we employ two algorithms, error effusion and direct binary search (DBS).
Two criteria for evaluating the performance of this process are: the visual
quality between the original image f[n,m] and the halftone f~[n,m] based
on human visual system (HVS) error metric, and the similarity between the
watermarked image g[n,m] and scanned back image g~[n,m] based on their
correlation. Finally, a cost function to jointly maximize the above
criteria is evaluated. By adopting these metrics, our project is able to
jointly maximize watermarking and halftoning against grayscale images, and
resilient to printing and scanning.


11 6:40-7:00
Transform domain steganography and adaptive steganography - A comparison
Vishnu Prakash Mani, Rajarajan Palanimurugan, Neerad Phansalkar, Siddharth Sundar
:
Steganography is the science of communicating secret data in an appropriate
multimedia carrier, e.g., image, audio, video and text file. The goal of
steganography is to conceal the very existence of the embedded data. Image
based steganography methods can be classified into three types: spatial domain
methods, transform domain methods and adaptive methods. The focus of this
project is to compare the performance of transform domain and adaptive
methods. For transform domain method, we consider a DWT based method in which
the secret data are embedded in the DWT coefficients of the carrier image. For
the adaptive method we consider a method of embedding data in the skin regions
of the carrier image. This requires the detection of skin regions in the
carrier image and then the secret data are embedded in the DWT coefficients of
these regions. These two methods will then be evaluated based on performance
criteria such as the amount of data that can be embedded in the carrier image,
the amount of distortion of the cover image (measured by PSNR), and the
robustness to transformations (e.g. image rotation, translation, cropping,
compression etc.). We Expect the adaptive method to be more robust than the
transform domain method but at the price of reduced data embedding capacity.


** 7:10 awards