Turbulence at the edge of the plasma in a nuclear fusion reactor can cause loss of confinement of the plasma. In
an effort to study the edge turbulence, the National Spherical Torus Experiment uses a gas puff imaging (GPI)
diagnostic to capture images of the turbulence. A gas puff is injected into the torus and visible light emission
from the gas cloud is captured by an ultra high-speed camera. Our goal is to detect and track coherent structures
in the GPI images to improve our understanding of plasma edge turbulence. In this paper, we present results
from various segmentation methods for the identification of the coherent structures. We consider three categories
of methods - immersion-based, region-growing, and model-based - and empirically evaluate their performance on
four sample sequences. Our preliminary results indicate that while some methods can be sensitive to the settings
of parameters, others show promise in being able to detect the coherent structures.
The detection of moving objects in complex scenes is the basis of many applications in surveillance, event
detection, and tracking. Complex scenes are difficult to analyze due to camera noise and lighting conditions.
Currently, moving objects are detected primarily using background subtraction algorithms, with block matching
techniques as an alternative. In this paper, we complement our earlier work on the comparison of background
subtraction methods by performing a similar study of block matching techniques. Block matching techniques
first divide a frame of a video into blocks and then determine where each block has moved from in the preceding
frame. These techniques are composed of three main components: block determination, which specifies the
blocks; search methods, which specify where to look for a match; and, the matching criteria, which determine
when a good match has been found. In our study, we compare various options for each component using publicly
available video sequences of a traffic intersection taken under different traffic and weather conditions. Our results
indicate that a simple block determination approach is significantly faster with minimum performance reduction,
the three step search method detects more moving objects, and the mean-squared-difference matching criteria
provides the best performance overall.
Detection and tracking of moving objects is important in the analysis of video data. One approach is to maintain a background model of the scene and subtract it from each frame to detect the moving objects which can then be tracked using Kalman or particle filters. In this paper, we consider simple techniques based on salient points to identify moving objects which are tracked using motion correspondence. We focus on video with a large field of view, such as a traffic intersection with several buildings nearby. Such
scenes can contain several salient points, not all of which move between frames. Using public domain video and two types of salient points, we consider how to make these techniques computationally efficient for detection and tracking. Our early results indicate that salient regions obtained using the Lowe keypoints algorithm and the Scale-Saliency algorithm can be used successfully to track vehicles in moderate resolution video.
Shape and texture features have been used for some time for pattern recognition in datasets such as remote sensed imagery, medical imagery, photographs, etc. In this paper, we investigate shape and texture features for pattern recognition in simulation data. In particular, we explore which features are suitable for characterizing regions of interest in images resulting from fluid mixing simulations. Three texture features -- gray level co-occurrence matrices, wavelets, and Gabor filters -- and two shape features -- geometric moments and the angular radial transform -- are compared. The features are evaluated using a similarity retrieval framework. Our preliminary results indicate that Gabor filters perform the best among the texture features and the angular radial transform performs the best among the shape features. The feature which performs the best overall is dependent on how the groundtruth dataset is created.
Comparing the output of a physics simulation with an experiment is
often done by visually comparing the two outputs. In order to
determine which simulation is a closer match to the experiment, more
quantitative measures are needed. This paper describes our early
experiences with this problem by considering the slightly simpler
problem of finding objects in a image that are similar to a given
query object. Focusing on a dataset from a fluid mixing problem, we
report on our experiments using classification techniques from machine
learning to retrieve the objects of interest in the simulation data.
The early results reported in this paper suggest that machine learning
techniques can retrieve more objects that are similar to the query
than distance-based similarity methods.
Texture features have long been used in remote sensing applications to represent and retrieve image regions similar to a query region. Various representations of texture have been proposed based on the Fourier power spectrum, spatial co-occurrence, wavelets, Gabor filters, etc. These representations vary in their computational complexity and their suitability for representing different region types. Much of the work done thus far has focused on panchromatic imagery at low to moderate spatial resolutions, such as images from Landsat 1-7 which have a resolution of 15-30 m/pixel, and from SPOT 1-5 which have a resolution of 2.5-20 m/pixel. However, it is not clear which texture representation works best for the new classes of high resolution panchromatic (60-100 cm/pixel) and multi-spectral (4 bands for red, green, blue, and near infra-red at 2.4-4 m/pixel) imagery. It is also not clear how the different spectral bands should be combined. In this paper, we investigate the retrieval performance of several different texture representations using multi-spectral satellite images from IKONOS. A query-by-example framework, along with a manually chosen ground truth dataset, allows different combinations of texture representations and spectral bands to be compared. We focus on the specific problem of retrieving inhabited regions from images of urban and rural scenes. Preliminary results show that 1) the use of all spectral bands improves the retrieval performance, and 2) co-occurrence, wavelet and Gabor texture features perform comparably.
Identifying moving objects from a video sequence is a fundamental and
critical task in many computer-vision applications. A common approach
is to perform background subtraction, which identifies moving objects
from the portion of a video frame that differs significantly from a
background model. There are many challenges in developing a good
background subtraction algorithm. First, it must be robust against
changes in illumination. Second, it should avoid detecting
non-stationary background objects such as swinging leaves, rain, snow,
and shadow cast by moving objects. Finally, its internal background
model should react quickly to changes in background such as starting
and stopping of vehicles. In this paper, we compare various background subtraction algorithms for detecting moving vehicles and pedestrians in urban traffic video sequences. We consider approaches varying from simple techniques such as frame differencing and adaptive median filtering, to more sophisticated probabilistic modeling techniques. While complicated techniques often produce superior performance, our experiments show that simple techniques such as adaptive median filtering can produce good results with much lower computational complexity.
The use of partial differential equations in image processing has become an active area of research in the last few years. In particular, active contours are being used for image segmentation, either explicitly as snakes, or implicitly through the level set approach. In this paper, we consider the use of the implicit active contour approach for segmenting scientific images of pollen grains obtained using a scanning electron microscope. Our goal is to better understand the pros and cons of these techniques and to compare them with the traditional approaches such as the Canny and SUSAN edge detectors. The preliminary results of our study show that the level set method is computationally expensive and requires the setting of several different parameters. However, it results in closed
contours, which may be useful in separating objects from the background in an image.
The automated production of maps of human settlement from recent satellite images is essential to detailed studies of urbanization, population movement, and the like. Commercial satellite imagery is becoming available with sufficient spectral and spatial resolution to apply computer vision techniques previously considered only for laboratory (high resolution, low noise) images. In this paper we attempt to extract human settlement from IKONOS 4-band and panchromatic images using spectral segmentation together with a form of generalized second-order statistics and detection of edges, corners, and other candidate human-made features in the imagery.
The automated production of maps of human settlement from recent satellite images is essential to studies of urbanization, population movement, and the like. The spectral and spatial resolution of such imagery is often high enough to successfully apply computer vision techniques. However, vast amounts of data have to be processed quickly. In this paper, we propose an approach that processes the data in several different stages. At each stage, using features appropriate to that stage, we identify the portion of the data likely to contain information relevant to the identification of human settlements. This data is used as input to the next stage of processing. Since the size of the data has reduced, we can now use more complex features in this next stage. These features can be more
representative of human settlements, and also more time consuming to
extract from the image data. Such a hierarchical approach enables us to process large amounts of data in a reasonable time, while maintaining the accuracy of human settlement identification. We illustrate our multi-stage approach using IKONOS 4-band and panchromatic images, and compare it with the straight-forward processing of the entire image.
PDE-based, non-linear diffusion techniques are an effective way to denoise images.In a previous study, we investigated the effects of different parameters in the implementation of isotropic, non-linear diffusion. Using synthetic and real images, we showed that for images corrupted with additive Gaussian noise, such methods are quite effective, leading to lower mean-squared-error values in comparison with spatial filters and wavelet-based approaches. In this paper, we extend this work to include anisotropic diffusion, where the diffusivity is a tensor valued function which can be adapted to local edge orientation. This allows smoothing along the edges, but not perpendicular to it. We consider several anisotropic diffusivity functions as well as approaches for discretizing the diffusion operator that minimize the mesh orientation effects. We investigate how these tensor-valued diffusivity functions compare in image quality, ease of use, and computational costs relative to simple spatial filters, the more complex bilateral filters, wavelet-based methods, and isotropic non-linear diffusion based techniques.
Detecting and tracking objects in spatio-temporal datasets is an
active research area with applications in many domains. A common
approach is to segment the 2D frames in order to separate the objects
of interest from the background, then estimate the motion of the
objects and track them over time. Most existing algorithms assume
that the objects to be tracked are rigid. In many scientific
simulations, however, the objects of interest evolve over time and
thus pose additional challenges for the segmentation and tracking
tasks. We investigate efficient segmentation methods in the context of
scientific simulation data. Instead of segmenting each frame
separately, we propose an incremental approach which incorporates the
segmentation result from the previous time frame when segmenting the
data at the current time frame. We start with the simple K-means
method, then we study more complicated segmentation techniques based
on Markov random fields. We compare the incremental methods to the
corresponding sequential ones both in terms of the quality of the
results, as well as computational complexity.
Proc. SPIE. 5102, Independent Component Analyses, Wavelets, and Neural Networks
KEYWORDS: Independent component analysis, Climatology, Data modeling, Principal component analysis, Solids, Electroluminescence, Data centers, Atmospheric modeling, Signal analyzers, Statistical analysis
Observed and simulated global temperature series include the effects
of many different sources, such as volcano eruptions and El Nino
Southern Oscillation (ENSO) variations. In order to compare the
results of different models to each other, and to the observed data,
it is necessary to first remove contributions from sources that are
not commonly shared across the models considered. Such a separation
of sources is also desired in order to assess the effect of human
contributions on the global climate. Atmospheric scientists currently use parametric models and iterative techniques to remove the effects of volcano eruptions and ENSO variations from global temperature trends. Drawbacks of the parametric approach include the non-robustness of the results to the estimated values of the parameters, and the possible lack of fit of the data to the model. In this paper, we investigate ICA as an alternative method for separating independent sources in global temperature series. Instead of fitting parametric models, we let the data guide the estimation, and separate automatically the effects of the underlying sources. We first assess ICA on simple artificial datasets to establish the conditions under which ICA is feasible in our context, then we study its results on climate data from the National Centers for Environmental Predictions.
In this paper, a novel data mining approach to address damage detection within the large-scale complex structures is proposed. Every structure is defined by the set of finite elements that also represent the number of target variables. Since large-scale complex structures may have extremely large number of elements, predicting the failure in every single element using the original set of natural frequencies as features is exceptionally time-consuming task. Therefore, in order to reduce the time complexity we propose a hierarchical localized approach for partitioning the entire structure into substructures and predicting the failure within these substructures. Unlike our previous sub-structuring approach, which is based on physical substructures in the structure, here we propose to partition the structure into sub-structures employing hierarchical clustering algorithm that also allows localizing the damage in the structure. Finally, when the identified substructure with a failure consists of sufficiently small number of target variables the extent of the damage in the element of the substructure is predicted. A numerical example analyses on an electric transmission tower frame is presented to demonstrate the effectiveness of the proposed method.
Removing noise from data is often the first step in data analysis. Denoising techniques should not only reduce the noise, but do so without blurring or changing the location of the edges. Many approaches have been proposed to accomplish this; in this paper, we focus on one such approach, namely the use of non-linear diffusion operators. This approach has been studied extensively from a theoretical viewpoint ever since the 1987 work of Perona and Malik showed that non-linear filters outperformed the more traditional linear Canny edge detector. We complement this theoretical work by investigating the performance of several isotropic diffusion operators on test images from scientific domains. We explore the effects of various parameters such as the choice of diffusivity function, explicit and implicit methods for the discretization of the PDE, and approaches for the spatial discretization of the non-linear operator etc. We also compare these schemes with simple spatial filters and the more complex wavelet-based shrinkage techniques. Our empirical results show that, with an appropriate choice of parameters, diffusion-based schemes can be as effective as competitive techniques.
In this paper, we describe the use of data mining techniques to search for radio-emitting galaxies with a bent-double morphology. In the past, astronomers from the FIRST (Faint Images of the Radio Sky at Twenty-cm) survey identified these galaxies through visual inspection. This was not only subjective but also tedious as the on-going survey now covers 8000 square degrees, with each square degree containing about 90 galaxies. In this paper, we describe how data mining can be used to automate the identification of these galaxies. We discuss the challenges faced in defining meaningful features that represent the shape of a galaxy and our experiences with ensembles of decision trees for the classification of bent-double galaxies.
Decision tress have long been popular in classification as they use simple and easy-to-understand tests at each node. Most variants of decision trees test a single attribute at a node, leading to axis- parallel trees, where the test results in a hyperplane which is parallel to one of the dimensions in the attribute space. These trees can be rather large and inaccurate in cases where the concept to be learned is best approximated by oblique hyperplanes. In such cases, it may be more appropriate to use an oblique decision tree, where the decision at each node is a linear combination of the attributes. Oblique decision trees have not gained wide popularity in part due to the complexity of constructing good oblique splits and the tendency of existing splitting algorithms to get stuck in local minima. Several alternatives have been proposed to handle these problems including randomization in conjunction wiht deterministic hill-climbing and the use of simulated annealing. In this paper, we use evolutionary algorithms (EAs) to determine the split. EAs are well suited for this problem because of their global search properties, their tolerance to noisy fitness evaluations, and their scalability to large dimensional search spaces. We demonstrate our technique on a synthetic data set, and then we apply it to a practical problem from astronomy, namely, the classification of galaxies with a bent-double morphology. In addition, we describe our experiences with several split evaluation criteria. Our results suggest that, in some cases, the evolutionary approach is faster and more accurate than existing oblique decision tree algorithms. However, for our astronomical data, the accuracy is not significantly different than the axis-parallel trees.
Advances in technology have enabled us to collect data from observations, experiments, and simulations at an ever increasing pace. As these data sets approach the terabyte and petabyte range, scientists are increasingly using semi-automated techniques from data mining and pattern recognition to find useful information in the data. In order for data mining to be successful, the raw data must first be processed into a form suitable for the detection of patterns. When the data is in the form of images, this can involve a substantial amount of processing on very large data sets. To help make this task more efficient, we are designing and implementing an object-oriented image processing toolkit that specifically targets massively-parallel, distributed-memory architectures. We first show that it is possible to use object-oriented technology to effectively address the diverse needs of image applications. Next, we describe how we abstract out the similarities in image processing algorithms to enable re-use in our software. We will also discuss the difficulties encountered in parallelizing image algorithms on the massively parallel machines as well as the bottlenecks to high performance. We will demonstrate our work using images from an astronomical data set, and illustrate how techniques such as filters and denoising through the thresholding of wavelet coefficients can be applied when a large image is distributed across several processors.