Computer simulation of metal alloys is an emerging trend in materials development. Simulated replicas of
fabricated alloys are based on the segmentations of alloy micrographs. Therefore, accurate segmentation of visible
precipitates is paramount to simulation accuracy. Since the shape and size of precipitates are key indicators of
physical alloy properties, automated segmentation algorithms must account for abundant prior information of
precipitate shape. We present a new method for constructing a prior enforcing rectangular shape which can be
applied within a min-cut framework for maximum a-posteriori segmentation.
Pose estimation and tracking of articulated objects like humans is particularly difficult due to the complex
occlusions among the articulated parts. Without the benefit of multiple views, resolution of occlusions becomes
both increasingly valuable and challenging. We propose a method for articulated 3D pose estimation from
monocular video which uses nonparametric belief propagation and employs a novel and efficient approach to
occlusion reasoning. We present a human tracking application, and evaluate results using the HumanEva II data
KEYWORDS: Signal to noise ratio, Detection and tracking algorithms, Data modeling, Scanners, Image restoration, Hemodynamics, Algorithm development, Performance modeling, Functional magnetic resonance imaging, Brain
The objective of fMRI data analysis is to detect the region of the brain that gets activated in response to a specific
stimulus presented to the subject. We develop a new algorithm for activation detection in event-related fMRI
data. We utilize a forward model for fMRI data acquisition which explicitly incorporates physiological noise,
scanner noise and the spatial blurring introduced by the scanner. After slice-by-slice image restoration procedure
that independently restores each data slice corresponding to each time index, we estimate the parameters of the
hemodynamic response function (HRF) model for each pixel of the restored data. In order to enforce spatial
regularity in our estimates, we model the prior distribution of the HRF parameters as a generalized Gaussian
Markov random field (GGMRF) model. We develop an algorithm to compute the maximum a posteriori (MAP)
estimates of the parameters. We then threshold the amplitude parameters to obtain the final activation map. We
illustrate our algorithm by comparing it with the widely used general linear model (GLM) method. In synthetic
data experiments, under the same probability of false alarm, the probability of correct detection for our method
is up to 15% higher than GLM. In real data experiments, through anatomical analysis and benchmark testing
using block paradigm results, we demonstrate that our algorithm produces fewer false alarms than GLM.
We develop real-time, low-complexity image classification algorithms suitable for a copy mode selector embedded in a low-end copier. The algorithms classify scanned images represented in RGB or in an opponent color space. Classes are the eight combinations of mono/color and text/mix/picture/photo. Classification is 30–98% accurate with misclassifications tending to be benign. The algorithms provide for improved copy quality, a simplified user interface, and increased copy rate.
We apply stabilized inverse diffusion equations (SIDEs) to segment microscopy images of materials to aid in
analysis of defects. We extend SIDE segmentation methods and demonstrate the effectiveness of our approaches
to two material analysis tasks. We first develop a method to successfully isolate the textured area of a solidification
defect to pixel accuracy. The second task involves utilizing multiple illuminations of the same structure of a
polycrystalline alloy. Our novel approach features the fusion of data extracted from each of these images to
create a composite segmentation which effectively represents all texture boundaries visible in any of the images.
These two methods both propose new techniques to incorporate multiple images to produce segmentations.
We propose a stochastic grammar model for random-walk-like time series that has features at several temporal scales.
We use a tree structure to model these multiscale features. The inside-outside algorithm is used to estimate the model
parameters. We develop an algorithm to forecast the sign of the first difference of a time series. We illustrate the algorithm
using log-price series of several stocks and compare with linear prediction and a neural network approach. We furthermore
illustrate our algorithm using synthetic data and show that it significantly outperforms both the linear predictor and the
neural network. The construction of our synthetic data indicates what types of signals our algorithm is well suited for.
We construct a hierarchical image grammar model based on stochastic grammars and apply it to document images. An efficient maximum a posteriori probability estimation algorithm for this model produces accurate segmentations of document images and classifications of image parts.
This paper reviews recent best basis search algorithms. The problem under consideration is to select a representation from a dictionary which minimizes an additive cost function for a given signal. We describe a new framework of multitree dictionaries, and an efficient algorithm for finding the best representation in a multitree dictionary. We illustrate the algorithm through image compression examples.
We develop a new methodology for constructing hierarchical stochastic image models called spatial random trees (SRTs) which admit polynomial-complexity exact inference algorithms. We use our framework of multitree dictionaries as the starting point for this construction. We develop an efficient algorithm for computing the EM updates and use it to estimate the model parameters. We illustrate our models and algorithms through image classification experiments.
We present a multiscale region-merging segmentation algorithm based on nonlinear diffusion equations. The algorithm is applicable to vector-valued images such as color images, or feature images obtained by pre-processing a texture image. The algorithm is experimentally shown to provide accurate segmentations for texture images.
We propose new best basis search algorithms for local cosine dictionaries. We provide several algorithms for dictionaries of various complexity. Our framework generalizes the classical best local cosine basis selection based on a dyadic tree.
SC964: HD Photo/JPEG XR in the Context of Modern Image Compression
The first part of this short course is a general introduction to image compression. We describe the basic structure of an image encoder and discuss its desirable properties. We describe various transforms that have led to various image compression algorithms such as JPEG, JPEG2000 and HD Photo/JPEG XR---specifically, block DCT, wavelets, and lapped transforms. The second part of the tutorial is a more in-depth look at the HD Photo compression algorithm (recently standardized as JPEG XR) developed by Microsoft Corporation. We describe various aspects of the algorithm and, whenever possible, discuss its modules in the context of the corresponding modules of the JPEG compression standard. We conclude by presenting a comparative performance analysis of HD Photo, JPEG2000, SPIHT and JPEG, using both PSNR and the perceptual SSIM index as distortion metrics.
(Please define the acronyms SPIHT, PSNR, SSIM, and any other acronyms you may use in the Learning Outcomes or other areas of the description.)