- Prospective Students
- Undergraduate Programs
- Graduate Programs
April 25, 2018 at 12:00pm - Ph.D. Thesis Defense - RONALD M. KEMKER - LOW-SHOT LEARNING FOR THE SEMANTIC SEGMENTATION OF REMOTE SENSING IMAGERY
DIRS Laboratory 76-3215
April 25, 2018 at 12:00pm
LOW-SHOT LEARNING FOR THE SEMANTIC SEGMENTATION OF REMOTE SENSING IMAGERY
RONALD M. KEMKER
Ph.D. Thesis Defense
Deep-learning frameworks have made remarkable progress thanks to the creation of large annotated datasets such as ImageNet, which has over one million training images. Although this works well for color (RGB) imagery, labeled datasets for other sensor modalities (e.g., multispectral and hyperspectral) are minuscule in comparison. This is because annotated datasets are expensive and man-power intensive to complete; and since this would be impractical to accomplish for each type of sensor, current state-of-the-art approaches in computer vision are not ideal for remote sensing problems. The shortage of annotated remote sensing imagery beyond the visual spectrum has forced researchers to embrace unsupervised feature extracting frameworks. These features are learned on a per-image basis, so they tend to not generalize well across other datasets. In this dissertation, we propose three new strategies for learning feature extracting frameworks with only a small quantity of annotated image data; including 1) self-taught feature learning, 2) domain adaptation with synthetic imagery, and 3) semi-supervised classification. ``Self-taught'' feature learning frameworks are trained with large quantities of unlabeled imagery, and then these networks extract spatial-spectral features from annotated data for supervised classification. Synthetic remote sensing imagery can be used to boot-strap a deep convolutional neural network, and then we can fine-tune the network with real imagery. Semi-supervised classifiers prevent overfitting by jointly optimizing the supervised classification task along side one or more unsupervised learning tasks (i.e., reconstruction). Although obtaining large quantities of annotated image data would be ideal, our work shows that we can make due with less cost-prohibitive methods which are more practical to the end-user.
April 25, 2018 at 3:00am - MS Thesis Defense - NAJAT A. ALHARBI - ALKALI ACTIVATED SLAG CHARACTERIZATION BY SCANNING ELECTRON MICROSCOPY AND X-RAY MICROSANALYSIS
April 25, 2018 at 3:00am
ALKALI ACTIVATED SLAG CHARACTERIZATION BY SCANNING ELECTRON MICROSCOPY AND X-RAY MICROSANALYSIS
NAJAT A. ALHARBI
MS Thesis Defense
Blast furnace slag is a non-metallic byproduct generated by the production of iron and steel in a blast furnace at temperatures in the range of 1400°-1600° C. The alkali activation of blast furnace slag has the potential to reduce the environmental impact of cementitious materials and to be applied in geographic zones where weather is a factor that negatively affects performance of materials based on Ordinary Portland Cement. Alkali-activated blast furnace slag cements have been studied since the 1930s due to its high compressive strength; they can exceed 100 MPa in 28 days. The low Ca/Si ratio in slag improves its resistance to aggressive chemical materials such as acids, chlorides and sulphates. Blast furnace slag is a highly heterogeneous material. It is well known that its chemical composition affects the physical properties of the alkali activated material, however there is little work on how these inhomogeneities affect the microstructure and pore formation. In this study we characterize slag cement activated with KOH using several methods: x-ray diffraction (XRD), transmission electron microscopy (TEM), scanning electron microscopy (SEM), x-ray microanalysis (EDS), and quantitative element mapping. Attention is focused on delineating the phases induced by the alkali activation, as these phases are important in determining the mechanical properties of the material. For the alkaline activated slag, we found four phases. One phase was the particles carried over from the unactivated slag, but with significant changes in the chemical composition. In addition, three other phases were found -- one is rich in hydrotalcite and two phases were calcium aluminum silicate hydrate (C-A-S-H) predominant..
February 21, 2017 at 2:00am - PhD Imaging Science Thesis Defense - KELLY LARABY - Landsat Land Surface Temperature Product: Global Validation and Uncertainty Estimation
DIRS Lab 76-3215
February 21, 2017 at 2:00am
Landsat Land Surface Temperature Product: Global Validation and Uncertainty Estimation
PhD Imaging Science Thesis Defense
Advisor: Dr. John Schott
Land surface temperature (LST) is an Earth system data record that is important to many areas of study such as change detection, climate research, and smaller scale applications such as monitoring lakes and farms. LST is often derived from satellite thermal imagery to achieve adequate spatial and temporal coverage. The Landsat series of satellites are an unparalleled and attractive choice for developing an LST product, because they provide the longest running source of continuously acquired multispectral imagery. Landsat also has moderate spatial and temporal resolutions, and its sensors and data archives are well calibrated. The land surface temperature can be derived from a single Landsat thermal band if the atmosphere and surface emissivity are well known for each scene. The primary function of our algorithm is to perform atmospheric compensation on a per-pixel level, but eventually our process will be integrated with a global emissivity database to form the full LST product.
The LST algorithm was initially limited to Landsat scenes in North America, which motivated our efforts to extend the algorithm’s operability to the entire globe. This effort allowed us to perform a thorough global validation for Landsat 7. Another portion of our work was focused on developing a method for estimating the uncertainty in the LST retrievals, so that users can make informed decisions on which pixels to use. This was accomplished by dividing the global validation data into different ranges of cloud proximity and transmission, then using the root mean square error (RMSE) for each group to help define uncertainty. When transmission was greater than 0.7 and clouds were at least 5 km away from the pixel of interest, the difference between our predictions and the observed error in LST had RMSEs of roughly 1 K. When a bias removal technique was used on the observed LST errors, the RMSEs for the same conditions were reduced to around 0.75 K. Based on these values, we are confident that our uncertainty estimation method will be a useful addition to the LST product.
October 12, 2016 at 8:00am - MS Thesis Defense - David Rhodes - Radiometrically Correct Synthetic Video Development of Thermal Vehicle Targets
October 12, 2016 at 8:00am
Radiometrically Correct Synthetic Video Development of Thermal Vehicle Targets
MS Thesis Defense
Collecting large scientific quality thermal infrared image and video data sets is an expensive time consuming endeavor. Thermal infrared imagers cost much more than comparable visible systems and require skilled experienced operators. Also, time and experienced personnel are required to collect quality ground truth. Often it is advantageous to perform computer simulations as an alternative to collecting image and video data with real camera systems. As long as enough physics is incorporated into the models to give accurately comparable results to real imagery, simulated data can be used interchangeably. Generating synthetic images and video has the added benefit of being flexible as the user has control over every aspect of the simulation. Simulations are not subject to restrictions such as location, weather conditions, time of day, or time of year. Ground truth is assigned instead of measured in the synthetic world so it is known a priori. This thesis illustrates a method of using the Digital Image and Remote Sensing Image Generation (DIRSIG) software to create simulated infrared images and video of validated thermal target vehicle models inside thermal infrared wide-area scenes. A finite difference heat propagation and surface temperature solver, ThermoAnalytics Multi-Service Electro-optic Signature (MuSES TM),was used to accurately model the emissive thermal target vehicles. Validation of the thermal target vehicle model was performed using images taken from a laboratory calibrated MWIR camera. Images taken with the calibrated camera of the same type of vehicle as the target model were compared to the synthetic images for the same conditions for validation. Target vehicle motion was added to the simulations through the use of Simulation of Urban Mobility (SUMO), DIRSIGs movement files, and custom python scripting. The output images from DIRSIG were then laced together into video. The resulting video was used to test three tracking algorithms illuminating each one’s strengths and weaknesses.
August 15, 2016 at 2:00am - M.S. Thesis Defense - Zhenlin Xu - 3D Subject-Atlas Image Registration for Micro-Computed Tomography Based Characterization of Drug Delivery in the Murine Cochlea
Carlson Bldg. (76) – Room 1275
August 15, 2016 at 2:00am
3D Subject-Atlas Image Registration for Micro-Computed Tomography Based Characterization of Drug Delivery in the Murine Cochlea
M.S. Thesis Defense
A wide variety of hearing problems can potentially be treated with local drug delivery systems capable of delivering drugs directly to the cochlea over extended periods of time. Developing and testing such systems requires accurate quantification of drug concentration over time. A variety of techniques have been proposed for both direct and indirect measurement of drug pharmacokinetics; direct techniques are invasive, whereas many indirect techniques are imprecise because they rely on assumptions about the relationship between physiological response and drug concentrations. One indirect technique, however, is capable of quantifying drug pharmacokinetics very precisely: Micro-Computed tomography (µCT) can provide a non-invasive way to measure the concentration of a contrast agent in the cochlea over time. In this thesis, we propose a systematic approach for analyzing µCT images to measure concentrations of the contrast agent ioversol in mouse cochlea. This approach requires segmenting and classifying the intra-cochlea structures from µCT images, which is done via 3D atlas-subject registration to a published atlas of the mouse cochlea. Labels of each intra-cochlear structure in the atlas are propagated through the registration transformation to the corresponding structures in the µCT images. Pixel intensities are extracted from three key intra-cochlea structures: scala tympani (ST), scala vestibuli (SV), scala media (SM) in the µCT images, and these intensities are mapped into concentrations using a linear model between solution concentration and image intensity that is determined in a previous calibration step. To localize this analysis, the ST, SV, SM are divided into several discrete components, and the concentrations are estimated in each component using a weighted average with weights determined by solving a nonhomogeneous Poisson equation with Dirichlet boundary conditions on the component boundaries. We illustrate this entire system on a series of µCT images of an anesthetized mouse that include a baseline scan (with no contrast agent) and a series of scans after injection of ioversol into the cochlea.
August 5, 2016 at 9:00am - Ph.D. Dissertation Defense - Troy R. McKay - Detection of Anomalous Vehicle Loading
Fish Bowl 76-1275
August 5, 2016 at 9:00am
Detection of Anomalous Vehicle Loading
Troy R. McKay
Ph.D. Dissertation Defense
Determining the mass of a vehicle from ground based passive sensor data is important for many security and traffic safety reasons. A vehicle consists of multiple dependent and independent systems that each respond differently to changes in vehicle mass. In some cases, the responses of these vehicle systems can be measured remotely. If these remotely sensed system responses are correlated to the vehicle's mass, and the required vehicle parameters were known, it would be possible to calculate the mass of the vehicle as a function of these responses.
The research described here investigates multiple vehicle phenomenologies and their correlation to vehicle load. Brake temperature, engine acoustics, exhaust output, tire temperature, tire deformation, vehicle induced ground vibration, suspension response, and engine torque induced frame twist were all evaluated and assessed as potential methods of remotely measuring a vehicle's mass. Extensive field experiments were designed and carried out using multiple sensors of various types; including microphones, accelerometers, high-speed video cameras, high-resolution video cameras, LiDAR, and thermal imagers. These experiments were executed at multiple locations and employed passenger vehicles, and commercial trucks with loads ranging from zero to beyond the recommended load capacity of the vehicle. The results of these experiments were used to determine if the signature for each phenomenology could be accurately observed remotely, and if so, how well they correlated to vehicle mass. The suspension response and engine torque induced frame twist phenomenologies were found to have the best correlation to vehicle mass of the phenomenologies considered, with correlation values of 90.5% and 97.7% respectively. Physics-based models were built for both the suspension response, and the engine torque induced frame twist phenomenologies. These models detailed the relationship between each phenomenology and the mass of the vehicle. Full-scale field testing was done using improved remote detection methods, and the results were used to validate the physics-based models. The results of the full-scale field testing showed that both phenomenology could accurately calculate the mass of the vehicle remotely, given that certain vehicle parameters were accurately known. The engine torque induced frame twist phenomenology was able to find the mass of the test vehicle to within 10% of the true mass. Using the suspension response phenomenology the mass was accurately predicted as a function of its location on the vehicle. For either phenomenology to be effective, certain vehicle parameters must be known accurately; specifically the spring constant and damping coefficients of the vehicle's suspension, the unloaded mass, the unloaded center of gravity, and the unloaded moment of inertia of the vehicle. The models were also used to propagate measurement and parameter uncertainty through the vehicle mass calculation to arrive at the uncertainty in the mass estimation. Finally, the results of both the phenomenologies were combined into a single vehicle mass estimate with a smaller uncertainty than the individual vehicle system estimations taken alone.
July 28, 2016 at 3:00am - Masters Thesis Defense - JOSEPH SVEJKOSKY - Hyperspectral Vehicle BRDF Learning: Seeking Illuminant Invariant Signatures For Vehicle Reacquisition and Tracking
DIRS Lab 76-3215
July 28, 2016 at 3:00am
Hyperspectral Vehicle BRDF Learning: Seeking Illuminant Invariant Signatures For Vehicle Reacquisition and Tracking
Masters Thesis Defense
The spectral signatures of vehicles in hyperspectral imagery exhibit temporal variations due to the preponderance of surfaces with material properties that display non-Lambertain bi-directional reflectance functions (BRDFs). These temporal variations are caused by changing illumination conditions, changing sun-target-sensor geometry, changing road surface properties, and changing vehicle orientations. To quantify these variations and determine the relative importance of each in a vehicle reacquisition and tracking scenario, a hyperspectral vehicle BRDF sampling experiment was conducted in which four vehicles were rotated at different orientations and imaged over a six-hour period. The results illustrate the need for a target model and detection scheme that incorporate non-Lambertian BRDFs. The proposed model seeks to learn a vehicle BRDF from a series of images and then apply the learned BRDF for increased detection/reacquisition accuracy. This detection scheme is compared to sub space detections algorithms and graph-based detection algorithms in which the target BRDF is not accounted for. The algorithms are compared using a test environment in which observed spectral signatures from the experiments are implanted into aerial hyperspectral imagery of a similar GSD that contain large quantities of vehicles.
July 26, 2016 at 9:00am - PhD Thesis Defense - Can Jin - Characterization and Reduction of Noise in Manifold Representations of Hyperspectral Imagery
Fish Bowl 76-1275
July 26, 2016 at 9:00am
Characterization and Reduction of Noise in Manifold Representations of Hyperspectral Imagery
PhD Thesis Defense
A new workflow to produce dimensionality reduced manifold coordinates based on the improvements of landmark Isometric Mapping (ISOMAP) algorithms using local spectral models is proposed. Manifold space from nonlinear dimensionality reduction better addresses the nonlinearity of the hyperspectral data and often has better performance comparing to the results of linear methods such as Minimum Noise Fraction (MNF). The dissertation mainly focuses on using adaptive local spectral models to further improve the performance of ISOMAP algorithms by addressing local noise issues and perform guided landmark selection and nearest neighborhood construction in local spectral subsets. This work could benefit the performance of common hyperspectral image analysis tasks, such as classification, target detection, etc., but also keep the computational burden low. This work is based on and improves the previous ENH-ISOMAP algorithm in various fronts. The workflow is based on a unified local spectral subsetting framework. The theory of embedding spaces in local spectral subsets can serve as local noise models is first proposed and used to perform noise estimation, MNF regression and guided landmark selection in a local sense. Passive and active methods are proposed and verified to select landmarks deliberately to ensure local geometric structure coverage and local noise avoidance. Then, a novel local spectral adaptive method is used to construct k-nearest neighbor graph. Finally, a post-MNF transformation in the manifold space is also introduced to further compress the signal dimensions. The workflow is implemented using C++ with multiple implementation optimizations, including using heterogeneous computing platforms that are available in personal computers. The results are presented and evaluated by Jeffrey-Matsushita separability metric, as well as the classification accuracy of supervised classifiers. The proposed workflow shows significant and stable improvements over the dimensionality reduction performance from traditional MNF and ENH-ISOMAP on various hyperspectral datasets. The running speed of the proposed implementation is also improved.
July 19, 2016 at 10:00am - PhD.Thesis Defense - JUSTIN HARMS - The Design and Implementation of GRIT-T: RIT’s Next-generation Field-Portable Goniometer System
July 19, 2016 at 10:00am
The Design and Implementation of GRIT-T: RIT’s Next-generation Field-Portable Goniometer System
Various field portable goniometers have been designed to capture in-situ measurements of a materials bi-directional reflectance distribution function (BRDF), each with a specific scientific purpose in mind. [26, 32, 28, 8] The Rochester Institute of Technology’s (RIT) Chester F. Carlson Center for Imaging Science recently created a novel instrument incorporating a wide variety of features into one compact apparatus in order to obtain very high accuracy BRDFs of short vegetation and sediments, even in undesirable conditions and austere environments. This next generation system integrates a dual-view design using two VNIR/SWIR spectroradiometers to capture target reflected radiance, as well as incoming radiance, to provide for better optical accuracy when measuring in non-ideal atmospheric conditions or when background illumination effects are non-negligible. The new, fully automated device also features a laser range finder to construct a surface roughness model of the target being measured, which enables the user to include inclination information into BRDF post-processing and further allows for roughness effects to be studied for radiative transfer modeling. The highly portable design features automatic leveling, a precision-engineered frame, and a variable measurement plane that allow for BRDF measurements on rugged, un-level terrain while still maintaining true angular measurements with respect to the target, all without sacrificing measurement speed. Despite the expanded capabilities and dual sensor suite, the system weighs less than 75 kg, which allows for excellent mobility and data collection on silty clay or fine sand.
July 19, 2016 at 1:00am - PhD Thesis Defense - Philip Salvaggio - Image Quality Modeling and Optimization for Non-Conventional Aperture Imaging Systems
Fish Bowl 76-1275
July 19, 2016 at 1:00am
Image Quality Modeling and Optimization for Non-Conventional Aperture Imaging Systems
PhD Thesis Defense
The majority of image quality studies have been performed on systems with conventional aperture functions. These systems have straightforward aperture designs and well-understood behavior. Image quality for these systems can be predicted by the General Image Quality Equation (GIQE). However, in order to continue pushing the boundaries of imaging, more control over the point spread function of an imaging system may be necessary. This requires modifications in the pupil plane of a system, causing a departure from the realm of most image quality studies. Examples include sparse apertures, synthetic apertures, coded apertures and phase elements. This work will focus on sparse aperture telescopes and the image quality issues associated with them, however, the methods presented will be applicable to other non-conventional aperture systems.
In this research, an approach for modeling the image quality of non-conventional aperture systems will be introduced. While the modeling approach is based in previous work, a novel validation study will be performed, which accounts for the effects of both broadband illumination and wavefront error. One of the key image quality challenges for sparse apertures is post-processing ringing artifacts. These artifacts have been observed in modeled data, but a validation study will be performed to observe them in measured data and to compare them to model predictions. Once validated, the modeling approach will be used to perform a small set of design studies for sparse aperture systems, including spectral bandpass selection and aperture layout optimization.
July 14, 2016 at 10:00am - Master’s Thesis Defense - GRANT ANDERSON - An evaluation of the silicon spectral range for determination of nutrient content of grape vines
DIRS Lab 76-3215
July 14, 2016 at 10:00am
An evaluation of the silicon spectral range for determination of nutrient content of grape vines
Master’s Thesis Defense
The grape industry relies on in situ crop assessment to aid in the day-to-day and seasonal management of their crop. In the case of soil-plant chemistry interactions, there are six key nutrients of interest to viticulturists in the growing of wine grapes: nitrogen, potassium, phosphorous, magnesium, zinc, and boron. Traditional methods of determining the levels of these nutrients are through collection and chemical analysis of petiole samples from the grape vines themselves. In this study, however, we collected ground-level observations of the spectra of the grape vines using a hyperspectral spectroradiometer (0.4-2.5µm range; 1nm resampled spectral interval) at the same time that petioles samples were harvested. The data were collected for two different grape cultivars, both during bloom and veraison phenological stages to provide analytical variability, while also considering the impact of temporal/seasonal change. The data were interpolated to 1nm bandwidths, yielding a consistent 1nm spectral resolution before comparing it to the nutrient data collected. Spectral reflectance also was resampled to match the 10nm bands used by the Airborne Visible and Infrared Imaging Spectrometer (AVIRIS); this was done to assess the efficacy of nutrient modeling using a more standard, airborne system’s spectral resolution. Our analysis was limited to the silicon photodiode range to increase the utility of the approach for wavelength-specific cameras (via spectral filters) in a low cost unmanned aerial vehicle (UAV) platform. Five different approaches were tested to fit the data to the nutrient data. These were: a narrow-band Normalized Difference Index (NDI) approach using a standard linear fit, step-wise linear regression (SLR) using the silicon range of wavelengths, SLR using the NDI that correlated highly with the nutrient data, SLR using the 1st derivative of the reflectance spectra, and SLR using continuum-removed spectra, applied over the red trough (560-750nm) spectral region. For 1nm reflectance data, these methods generated models for nutrient modeling using between 2-10 wavelengths, and associated coefficients of determination values ranging between R2 = 0.74-0.86 across the six nutrients. In the case of the 10nm resampled spectral data, model fits ranged between R2 = 0.61-0.93 across the six nutrients, using 2-18 unique wavelength bands. These results bode well for eventual non-destructive, accurate and precise assessment of vineyard nutrient status through the use of UAVs.
July 14, 2016 at 9:00am - PhD Thesis Defense - Alexandra B. Artusio-Glimpse - The Realization and Study of Optical Wings
Fish Bowl 76-1275
July 14, 2016 at 9:00am
The Realization and Study of Optical Wings
Alexandra B. Artusio-Glimpse
PhD Thesis Defense
Consider the airfoil: a carefully designed structure capable of stable lift in a uniform air flow. It so happens that air pressure and radiation (light) pressure are similar phenomena because each transfer momentum to flow-disturbing objects. This, then, begs the question: does an optical analogue to the airfoil exist? Though an exceedingly small effect, scientists harness radiation pressure in a wide gamut of applications from micromanipulation of single biological particles to the propulsion of large spacecrafts called solar sails. We introduce a cambered, refractive rod that undergoes optical forces analogous to those seen in aerodynamics, and I call this analogue the optical wing. Flight characteristics of optical wings are determined by wing shape and material in a uniform radiation field. The lift force and axial torque are functions of the wing's angle of attack with stable and unstable orientations. These structures operate as intensity-dependent, parametrically driven oscillators and exhibit bistability when analyzed in an accelerating frame. Experiments on semi-buoyant wings in water found semicylindrically shaped, refractive microparticles traversed a laser beam and rotated to an illumination-dependent stable orientation. Preliminary tests aid in the development of a calibrated force measurement experiment to directly evaluate the optical forces and torque on these samples. A foundational study of the optical wing, this work contributes to future advancements of flight-by-light.
July 12, 2016 at 10:00am - PhD Thesis Defense - Rajagopalan Rengarajan - Evaluation of sensor, environment and operational factors impacting the use of multiple sensor constellations for long term resource monitoring
DIRS Lab 76-3215
July 12, 2016 at 10:00am
Evaluation of sensor, environment and operational factors impacting the use of multiple sensor constellations for long term resource monitoring
PhD Thesis Defense
Moderate resolution remote sensing data oﬀers the potential to monitor the long and short term trends in the condition of the Earth’s resources atﬁner spatial scales and over longer time periods. While improved calibration (radiometric and geometric), free access (Landsat, Sentinel, CBERS), and higher level products in reﬂectance units have made it easier for the science community to derive the biophysical parameters from these remotely sensed data, a number of issues still aﬀect the analysis of multi-temporal datasets. These are primarily due to sources that are inherent in the process of imaging from single or multiple sensors. Some of these undesired or uncompensated sources of variation include variation in the view angles, illumination angles, atmospheric eﬀects, and sensor eﬀects such as Relative Spectral Response (RSR) variation between diﬀerent sensors. The complex interaction of these sources of variation would make their study extremely diﬃcult if not impossible with real data, and therefore, a simulated analysis approach is used in this study.
A synthetic forest canopy is produced using the Digital Imaging and Remote Sensing Image Generation (DIRSIG) model and its measured BRDFs are modeled using the RossLi canopy BRDF model. The simulated BRDF matches the real data to within 2% of the reﬂectance in both the red and the NIR spectral bands. The BRDF modeling process is extended to model and characterize the defoliation of a forest, which is used in factor sensitivity studies to estimate the eﬀect of each factor for varying environment and sensor conditions. Finally, a factorial experiment is designed to understand the signiﬁcance of the sources of variation, and regression based analysis are performed to understand the relative importance of the factors. The design of experiment and the sensitivity analysis conclude that the atmospheric attenuation and variations due to the illumination angles are the dominant sources impacting the at-sensor radiance.
July 8, 2016 at 2:00am - PhD Thesis Defense - DOUGLAS MACDONALD - Modeling the Radar Return of Powerlines Using an Incremental Length Diffraction Coefficient Approach
DIRS Lab 76-3215
July 8, 2016 at 2:00am
Modeling the Radar Return of Powerlines Using an Incremental Length Diffraction Coefficient Approach
PhD Thesis Defense
A method for modeling the signal from cables and powerlines in Synthetic Aperture Radar (SAR) imagery is presented. A popular tool that uses the geometric optics approximation for modeling imagery for remote sensing applications across a wide range of modalities is the Digitial Imaging and Remote Sensing Image Generation (DIRSIG) tool. The drawback to using DIRSIG at longer wavelengths is it does not account for diffraction. Since the characteristic diameter of many communication cables and powerlines are on the order of the wavelength of the incident radiation, diffraction is the dominant mechanism by which the radiation gets scattered for these targets. Comparison of DIRSIG imagery to field data shows good scene-wide qualitative agreement as well as Rayleigh distributed noise in the amplitude data, as expected for coherent imaging with speckle. A closer inspection of the Radar Cross Sections of canonical targets such as trihedrals and dihedrals, however, shows DIRSIG consistently underestimated the scattered return, especially away from specular observation angles. Powerlines were not apparent in the simulated data. For modeling powerlines outside of DIRSIG using a standalone approach, an Incremental Length Diffraction Coefficient (ILDC) method was used. The Radar Cross Sections produced by this method were accurate to within the experimental uncertainty of measured anechoic data for both X and C-band frequencies across an 80o arc for 5 different target types and diameters. For field data in an actual X-band circular SAR collection, a mean error of 3.3% for data with a measurement uncertainty of 3.3% was obtained in the HH channel. For the VV channel, a mean error of 9.9% was obtained for data with a measurement uncertainty of 3.4%. This error is likely due to scattering from the grooves in helically wound powerlines, which violate the smooth cylinder assumption made by this research. Future work for improving this method would likely entail adding a far-field semi-open waveguide contribution to the 2D diffraction coefficient for TE polarized radiation. Incorporating 2nd order diffractions would also improve accuracy for multiple closely spaced powerlines.
July 7, 2016 at 10:00am - PhD Thesis Defense - Leidy P. Dorado-Munoz - Spectral Target Detection using Schroedinger Eigenmaps
July 7, 2016 at 10:00am
Spectral Target Detection using Schroedinger Eigenmaps
Leidy P. Dorado-Munoz
PhD Thesis Defense
Hyperspectral imagery (HSI) as the output of an optical remote sensing process reflects the information about properties of objects and materials on the earth's surface. Applications include environmental monitoring, meteorology, mapping, surveillance, etc. Many of these tasks include the finding and detection of specific objects, usually few or small that are surrounded by other materials or objects that clutter the scene and hide the relevant information. This target detection process has been boosted lately by the use of HSI. Non-linear transformation methods, mainly based on manifold learning algorithms, have shown a potential use in HSI transformation, dimensionality reduction and classification. The reduction processes includes the transformation of the image data to a new space where the relevant and hidden information is more easily revealed. These non-linear methods perform dimensionality reduction while preserving the local structure of the data, causing a minimal loss of relevant information.
One of these non-linear methods is the Schroedinger Eigenmaps (SE) algorithm, which is based on the well-known Laplacian Eigenmaps (LE), and it has been introduced as a technique for semi-supervised classification. Both algorithms, LE and SE, include the creation of an adjacency graph as a means to represent the spectral connectivity of the data set, and the eigendecomposition of a significant operator that for SE is the Schroedinger operator. The Schroedinger operator includes by definition a potential term that gives the option to encode prior information about the materials present in a scene, and steers the transformation in some convenient directions in order to cluster similar pixels together. The use of the SE algorithm is proposed in this thesis as a basis for a target detection methodology that does not require assuming any statistical or geometric models, but that boosts the separability between the class of interest and the other classes present in the image. This is performed by taking advantage of the privileged location that target pixels would have in the Schroedinger space. The proposed methodology does not just include the transformation of the data to a lower dimensional space but also includes the definition of a detector that capitalizes on the theory behind SE. In addition, a knowledge propagation scheme is used to combine spectral and spatial information as a means to propagate the “potential constraints” to nearby points connected via the adjacency graphs. The hope is that the propagation scheme helps to reinforce weak connections and improve the separability between most of the target pixels and the background.
May 17, 2016 at 10:00am - PhD Dissertation Defense - Fan Wang - Understanding high resolution aerial imagery using computer vision techniques
Bldg. (76) - Room 1275 (Fishbowl)
May 17, 2016 at 10:00am
Understanding high resolution aerial imagery using computer vision techniques
PhD Dissertation Defense
Computer vision could make important contribution to the analysis of remote sensing or aerial imagery. However, the resolution of early satellite imagery is not sufficient to provide useful spatial features. The situation is changing with the advent of very-high-spatial-resolution (VHR) imaging sensors. The change makes it possible to use computer vision techniques to analysis of man-made structures. Meanwhile, the development of multi-view imaging techniques will allow the generation of accurate point clouds as ancillary knowledge.
This dissertation aims at developing computer vision algorithms for the high resolution aerial imagery analysis in the context of application problems including debris detection, building detection and roof condition assessment. High resolution aerial imagery and point clouds are provided by Pictometry International for this study.
Debris detection is needed for effective debris removal and allocation of limited resources. Significant advances in aerial image acquisition have greatly enabled the possibilities for rapid and automated detection of debris. In this dissertation, a robust debris detection algorithm is proposed. Large scale aerial image is partitioned into homogeneous regions by interactive segmentation. Debris areas are identified based on extracted texture feature.
Robust building detection is an important part of high resolution aerial imagery understanding. This dissertation develops a 3D scene classification algorithm for building detection using point clouds derived from multi-view imagery. Point clouds are divided into point clusters using Euclidean Clustering. Individual point clusters are identified based on extracted spectral and 3D structure features.
The inspection of roof condition is an important step of damage claim processing in the insurance industry. Automated roof condition assessment from remotely sensed images is proposed in this dissertation. Texture classification and bag-of-words model are performed to assess the roof condition using features derived from the whole rooftop. However, considering the complexity of residential rooftop, a more sophisticated method is proposed to divide the task into two stages: 1) roof segmentation, followed by 2) classification of segmented roof regions.
Contributions of this study include the development of algorithms for debris detection using 2D images and building detection using 3D point clouds. For roof condition assessment, the solutions to this problem are explored in two directions: features derived from the whole rooftop and features extracted from each roof segments. Through our research, roof segmentation followed by segments classification is found to be a more promising method and the workflow processing is developed and tested. Since the methodology to solve these problems is focused on the design of hand-crafted features, unsupervised feature extraction techniques using deep learning should be explored in future work.
May 16, 2016 at 10:00am - Ph. D Thesis Defense - BURAK UZKENT - AERIAL VEHICLE TRACKING USING A MULTI-MODAL SENSOR
Carlson DIRS Lab 76-3215
May 16, 2016 at 10:00am
AERIAL VEHICLE TRACKING USING A MULTI-MODAL SENSOR
Ph. D Thesis Defense
Vehicle tracking from an aerial platform poses a number of unique challenges including the small number of pixels representing a vehicle, large camera motion, and parallax error. For these reasons, it is a more challenging task than traditional object tracking and is generally tackled through a number of different sensor modalities. Recently, the Wide Area Motion Imagery (WAMI) sensor platform has received considerable attention as it can provide higher resolution single band imagery in addition to its large area coverage. Despite these advantages, there is still not enough feature information and most WAMI systems struggle to persistently track vehicles. Additional modalities, such as spectral data, can be cruical in identifying objects even in low resolution scenes and advances in sensor technology is starting to make hyperspectral data acquisition at video frame rates possible. For this reason, a multi-modal optical sensor concept is considered in this thesis to improve tracking in adverse scenes.
The sensor considered is based on the Rochester Institute of Technology Multi- object Spectrometer, which is capable of collecting limited hyperspectral data at desired locations in addition to full-frame single band imagery. By acquiring hyperspectral data quickly, tracking can be achieved at reasonable frame rates which is crucial for tracking. More spectral samples can lead to a huge volume of data, so the relatively high cost of hyperspectral data acquisition and transmission need to be taken into account to design a realistic tracking system. By collecting and analyzing the extended (spectral) data only for the pixels of interest, we can address or avoid the unique challenges posed by aerial tracking. To accomplish this, we integrate limited hyperspectral data to improve measurement-to-track association. Also, a hyperspectral data based target detection method is presented to avoid the parallax effect and reduce the clutter density. Finally, the proposed system is evaluated on realistic, synthetic scenarios generated by the Digital Image and Remote Sensing Image Generation software.
April 28, 2016 at 9:00am - Ph. D Thesis Defense - SIYU ZHU - Text Detection in Natural Scenes and Technical Diagrams with Convolutional Feature Learning and Cascaded Classification
Carlson Bldg. (76) - Room 1275 (Fishbowl)
April 28, 2016 at 9:00am
Text Detection in Natural Scenes and Technical Diagrams with Convolutional Feature Learning and Cascaded Classification
Ph. D Thesis Defense
An enormous amount of digital images are being generated and stored every day. Understanding text in these images is an important challenge with large impacts for academic, industrial and domestic applications. Recent studies address the difficulty of separating text targets from noise and background, all of which vary greatly in natural scenes. To tackle this problem, we develop a text detection system to analyze and utilize visual information in a data driven, automatic and intelligent way.
The proposed method incorporates features learned from data, including patch-based coarse-to-fine detection (Text-Conv), connected component extraction using region growing, and graph-based word segmentation (Word-Graph). Text-Conv is a sliding window-based detector, with convolution masks learned using the Convolutional k-means algorithm (Coates et. al, 2011). Unlike convolutional neural networks (CNNs), a single vector/layer of convolution mask responses are used to classify patches. An initial coarse detection considers both local and neighboring patch responses, followed by refinement using varying aspect ratios and rotations for a smaller local detection window. Different levels of visual detail from ground truth are utilized in each step, first using constraints on bounding box intersections, and then a combination of bounding box and pixel intersections. Combining masks from different Convolutional k-means initializations, e.g., seeded using random vectors and then support vectors improves performance. The Word-Graph algorithm uses contextual information to improve word segmentation and prune false character detections based on visual features and spatial context. Our system obtains pixel, character, and word detection f-measures of 93.14%, 90.26%, and 86.77% respectively for the ICDAR 2015 Robust Reading Focused Scene Text dataset, out-performing state-of-the-art systems, and producing highly accurate text detection masks at the pixel level.
To investigate the utility of our feature learning approach for other image types, we perform tests on 8-bit greyscale USPTO patent drawing diagram images. An ensemble of Ada-Boost classifiers with different convolutional features (MetaBoost) is used to classify patches as text or background. The Tesseract OCR system is used to recognize characters in detected labels and enhance performance. With appropriate pre-processing and post-processing, f-measures of 82% for part label location, and 73% for valid part label locations and strings are obtained, which are the best obtained to-date for the USPTO patent diagram data set used in our experiments.
To sum up, an intelligent refinement of convolutional k-means-based feature learning and novel automatic classification methods are proposed for text detection, which obtain state-of-the-art results without the need for strong prior knowledge. Different ground truth representations along with features including edges, color, shape and spatial relationships are used coherently to improve accuracy. Different variations of feature learning are explored, e.g. support vector-seeded clustering and MetaBoost, with results suggesting that increased diversity in learned features benefit convolution-based text detectors.
April 6, 2016 at 1:00am - PhD Dissertation Defense - GARRETH RUANE - OPTIMAL PHASE MASKS FOR HIGH-CONTRAST IMAGING APPLICATIONS
April 6, 2016 at 1:00am
OPTIMAL PHASE MASKS FOR HIGH-CONTRAST IMAGING APPLICATIONS
PhD Dissertation Defense
Phase-only optical elements can provide a number of important functions for high-contrast imaging. This thesis presents analytical and numerical optical design methods for accomplishing specific tasks, the most significant of which is the precise suppression of light from a distant point source. Instruments designed for this purpose are known as coronagraphs. Here, advanced coronagraph designs are presented that offer improved theoretical performance in comparison to the current state-of-the-art. Applications of these systems include the direct imaging and characterization of exoplanets and circumstellar disks with high sensitivity. Several new coronagraph designs are introduced and, in some cases, experimental support is provided.
In addition, two novel high-contrast imaging applications are discussed: the measurement of sub-resolution information using coronagraphic optics and the protection of sensors from laser damage. The former is based on experimental measurements of the sensitivity of a coronagraph to source displacement. The latter discussion presents the current state of ongoing theoretical work. Beyond the mentioned applications, the main outcome of this thesis is a generalized theory for the design of optical systems with one of more phase masks that provide precise control of radiation over a large dynamic range, which is relevant in various high-contrast imaging scenarios. The optimal phase masks depend on the necessary tasks, the maximum number of optics, and application specific performance measures. The challenges and future prospects of this work are discussed in detail.
March 14, 2016 at 2:30am - PhD Dissertation Defense - SHRUTI GOPAL - INTER – SUBJECT VARIABILITY ANALYSIS IN BLIND SOURCE SEPARATION APPLICATIONS
Carlson DIRS Lab 76-3215
March 14, 2016 at 2:30am
INTER – SUBJECT VARIABILITY ANALYSIS IN BLIND SOURCE SEPARATION APPLICATIONS
PhD Dissertation Defense
The holy grail of brain imaging is the identification of a biomarker, which can identify an abnormality that can be used both, to diagnose disease and track the effectiveness of treatment and disease progression. Typically approaches that search for biomarkers start by identifying mean differences between groups of patients and healthy controls. However, combining data from different subjects and groups to be able to make meaningful inferences is not trivial. The structure of the brain in each individual is unique in the size and shape as well as in the relative location of anatomical landmarks (e.g. sulci and gyri). When looking for mean differences in functional images, this issue is exacerbated by the presence of variability in functional localization i.e. variability in the location of functional regions in the brain. This is notably an important reason to focus on looking for inter-individual differences or variability.
Inter-subject variability in neuroimaging experiments is often viewed as noise. The analyses are setup in a manner to ignore this variability assuming that a global spatial normalization brings the data into the same space. Nonetheless, functional activation patterns can be impacted by variability in multiple ways for e.g., there could be spatial variability of the maps or variability in the spectral composition of the timecourses or variability in the connectivity between the activation patterns identified. The overarching problem this thesis seeks to contribute to, is seeking improved measures to quantify spatial, spectral and connectivity based variability and to identify associated cognitive or behavioral differences in the distribution of brain networks. We have successfully shown that different (spatial and spectral) measures of variability in blind source separated functional activation patterns underline previously unexplained characteristics that help in discerning schizophrenia patients from healthy controls. Additionally, we show that variance measures in dynamic connectivity between networks in healthy controls can justify relationship between connectivity patterns and executive functioning abilities.