ECE Masters Theses

Energy Conservation for Content-based Image Retrieval on Mobile Devices

Karthik Kumar — Sat, 26 Jul 2008 11:19:39 PDT

Battery-powered mobile systems such as PDAs (personal digital assistants) and mobile phones play an increasing role in handling visual contents such as images. Thousands of images can be stored in a mobile system with the advances in memory technology; this creates the need for better organization and retrieval of the images. Content Based Image Retrieval (CBIR) provides a method to retrieve images based on their contents. In CBIR, images are represented and compared by high-dimensional vectors called features. Loading these features into memory and comparing them consumes a significant amount of energy. In this thesis, we present the first study on energy conservation for CBIR on a mobile system. We develop an adaptive loading scheme to save energy for CBIR. Our method reduces the features to be loaded into memory for each query image. The reduction is achieved by estimating the difficulty of the query. If the images are dissimilar, fewer features are sufficient; less computation is performed and energy can be saved. For each query image, a similarity index is calculated to determine the features' length for meeting a target accuracy. Further, we consider the effect of consecutive user queries and show how features can be "cached" in memory to save energy. We implemented this algorithm on an HP iPAQ hw6945 and measured the energy savings. For a collection of 5000 images, we obtained average energy reduction of 61.3% compared to an existing CBIR implementation.

Color Characterization and Modeling of a Scanner

Sanjyot A. Gindi — Fri, 25 Jul 2008 11:46:42 PDT

The quality of digital color image capture systems like scanners and digital cameras is determined by the range of colors they can accurately sense and record. This range of colors is known as the color gamut of the system. Each image capture system has its own color space defined by its set of illuminants and color sensors. Such a device-dependent color space is specific and unique to the system. In order that these capture systems be evaluated and compared, the colors sensed by them need to be plotted in a common color space which is independent of any particular system. Such color spaces, called device-independent color spaces, have been defined by Commission internationale de l'éclairage(CIE).

The goal of this research is to characterize and model a color scanner and determine its color gamut. Such a color gamut is defined by a transformation matrix that converts the colors in the device dependent color space to those in a standard device-independent color space like CIE XYZ, L*a*b*. In this work, two methods have been used to determine the transformation matrix, namely, a model-based method and a regression based method. The model based method describes the scanner system in terms of its lamp spectrum and the spectral responses of its sensors to the reflectance of the target to be scanned. The regression-based method finds a transformation matrix by mapping the color values of a particular target as recorded by the scanner to its actual values in XYZ space, by regression.

The accuracy with which each method describes the system has been determined and compared. In addition, the system under test is compared with other scanner systems by plotting the color gamuts of each.

Diesel Particulate Filter Diagnostics Using Correlation and Spectral Analysis

Pranati R. Surve — Fri, 25 Jul 2008 11:00:20 PDT

Diesel Particulate Filters (DPF) are used to trap the harmful particulate matter (PM) present in the exhaust of diesel engines. The particulate matter is trapped in and on a porous ceramic substrate to keep PM emissions low. The onboard diagnostics requirements enforced by Environmental Protection Agency (EPA) require that the DPF perform well to keep emissions below certain specified levels. Further, should the DPF fail in any way, resulting in higher emission levels, this event must be detected by the engine control module. The objective of this work is to “detect failed DPF condition”. The temperature and pressure signals from transducers inserted into the inlet and outlet of the DPF are analyzed. The approach is to correlate the pre-DPF and post-DPF temperature and pressure signals and define the transfer function characteristics for nominal DPF behavior. Determining how these characteristics change as a result of filter failure forms the basis of a DPF fault detection algorithm. It is observed from the test data that for the pressure signal, other than the mean value signal (i.e., at zero frequency), most of the energy content is concentrated at the firing frequency of the engine. The dynamic pressure signals are used to determine the magnitude squared of the transfer function characteristics of DPF by energy spectral analysis. This approach can achieve a failure detection of lightly failed DPF which is not possible by current algorithms based on mean value pressure drop. The most significant contribution of this research is the extension of dynamic pressure signal analysis from steady-state engine operation to transient operating conditions.

Online Error Detection for High Data Rate Distributed Applications

Ignacio Laguna — Tue, 08 Jul 2008 07:56:54 PDT

Distributed systems comprising interacting services need runtime error detection to catch errors arising from software bugs, hardware errors, or unexpected operating conditions. A significant class of detection systems performs detection at the application level, based on the state of the application. For example, rule-based systems match rules against the application’s state deduced by the detection system at runtime. Many large-scale distributed applications generate a high rate of messages which can overwhelm the capacity of the detection system. An approach to handle this is sampling, that is, processing only a fraction of the messages. However, this approach leads to non-determinism with respect to the detection system’s view of what state the application is in. This in turn leads to inaccuracies in matching state-based rules causing degradation in the quality of detection. In this work, we present an approach to select the messages to sample and process such that the non-determinism is minimized. Next, we present a Hidden Markov Model-based technique to probabilistically identify which application states are most likely so that the detection system can perform rule-based detection for only those states. We demonstrate the techniques in a detection system called Monitor applied to a Java-based three-tier online banking system. The techniques do not need application modifications or a priori application model, but do require knowledge of expected application behavior to come up with the rules. We empirically evaluate accuracy and precision of detection under different load conditions and compare our solution with two other state-of-the-art systems: Pinpoint and Convolution algorithm.

A Model for Population Firing in Neuronal Networks with Planar Lattice Topology

Charity A. Pettis — Tue, 01 Jul 2008 18:23:06 PDT

The creation, modulation, propagation, and termination of seizures are of great interest to neuroscientists as understanding the dynamics of these responses may aid the prevention of seizures. Three major factors aﬀecting epileptiform activity are intrinsic neuronal properties, synaptic properties and network connectivity. The computation required for the full scale simulation of neuronal networks is prohibitively expensive. To address this issue, we develop a simpler model for networks with planar lattice topology and “small world” characteristics. Using this model, one may examine the eﬀects of connectivity and intrinsic neuronal properties on population ﬁring, a measure of network activity, quickly. To verify that the model is acceptable, the population firing produced by this simplified model is compared with population firing obtained from simulations of full-scale networks of model neurons for various connectivity parameters. To check that the behaviors exhibited by the networks are not unique to networks of a specific neuron model, the network is simulated using two neuron models: the Hodgkin-Huxley model and the leaky integrate-and-ﬁre model. With reasonable assumptions required to account for the differing levels of detail between the various models, it was observed that the network simulations and the simplified model exhibit similar behaviors.

Embedding and Detecting Machine Readable Data in Images Printed by Electrophotographic Printers

Vivek Shah — Tue, 29 Apr 2008 07:19:44 PDT

Watermarking is a technique used to authenticate and distinguish original documents or images from the copied or forged ones. Many watermarking methods have been developed and are being used for printed documents and images. Though these current watermarking techniques can distinguish between original and forged documents, it is not possible to identify the source of the forged and counterfeit documents, knowledge of which is required in some cases to stop future forgeries. In this thesis we propose a machine readable watermarking technique for images, halftoned using clustered-dot halftoning technique and printed on electrophotographic printer, that can, not only identify forged documents and images, but also provide information about the printer used and date on which the document was printed.

First, we propose a method of embedding data in the image by shifting clustered-dot pair diagonally and horizontally. The range of gray level values within which data can be embedded, number of pixels by which dots should be shifted and location of shifted dots is decided based on experiments carried out.

Next, we propose a method for detecting data from the scanned image. Scanned image is preprocessed using various image processing algorithms and techniques before detection stage. The detection of embedded data is done from the knowledge of the centroids of each clustered-dot. Error control coding technique is also proposed to improve the accuracy of detection process.

Lastly we propose an embedding method using pulse width modulation to reduce the visual distortions in the image caused by embedding the data.

Application of Vehicle Detector Waveforms in Vehicle Re-Identification and Evaluating Detector Installation Performance

Virgil F. Totten — Tue, 22 Apr 2008 13:40:55 PDT

Vehicle detectors in arterial roads and highways have long been used by the transportation management community for the purposes of determining vehicle presence at stop bars and traffic volume at a single point in a traffic stream. Advanced analysis of vehicle detectors and the waveforms they produce has recently become an area of increased interest in the transportation community. In this thesis, several applications of the waveforms produced by vehicle detectors will be explored.

Presently there is limited testing performed to evaluate a detector installation. By using vehicle signatures captured from vehicle detector waveforms, it is possible perform a more robust evaluation of detector installations. Several metrics for determining installation performance are presented.

Determining the travel time of a traffic stream over a significant distance is a much sought after goal by the transportation community. Vehicle signatures provide the foundation for a correlation based method of re-identifying vehicles over significant distances using existing transportation management infrastructure. Through the application of this vehicular re-identification, along with selective filtering of the data, it becomes possible to accurately determine, and track, the travel time of a traffic stream.

A Context Aware Content Based Federated Access Control System for Healthcare Domain

Haseeb Amjad — Mon, 10 Dec 2007 10:33:58 PST

Legislation to create electronic healthcare records and provide electronic healthcare services requires the same level of privacy and disclosure regulations as are applicable to the current practices for paper based patient health records. Most of work in this area has been organization-oriented that deals with exchange of information among healthcare organizations (such as referrals). However, the requirements for ensuring security and privacy of information for online access and sharing of health records in a federated healthcare environment have not been adequately addressed. To address this problem, we have developed a context-aware content-based access control policy specification framework, known as Generalized Temporal Role Based Access Control model (X-GTRBAC). This framework has been prototyped using an XML-driven federated environment that is comprised of synthetic healthcare multimedia databases. Our system integrates both privacy and disclosure policies with well-known healthcare standards used in the industry in order to specify the precise requirements of a practical healthcare system. In particular, the prototype uses the Clinical Document Architecture (CDA) of the Health Level 7 (HL7) organization as the underlying information model, and provides a methodology for associating user and environmental context parameters, with HL7 Reference Information Model [1]. The X-GTRBAC specification language used in this prototype can be broadly applied to a wide range of distributed and disparate healthcare applications for fine grained and flexible context-aware control for HL7 objects.

Intra-level Incomplete Bypassing: Achieving Performance and Power Efficiency

Eric P. Villasenor — Fri, 07 Dec 2007 09:59:20 PST

Researchers have proposed clustered microarchitectures to capture the benefits of high performance and high energy efficiency. Typically, clustered microarchitectures offer fast local bypasses (i.e., value forwarding between instructions) within clusters and require global bypasses to take longer, more than one cycle. With communication locality (i.e., most communication is within the clusters) the clustered designs capture the benefits of both improved instructions per cycle and increased clock-frequency. Traditional clustered microarchitectures are implemented by partitioning the register file and associated functional units to clusters. In this work, an alternate technique is demonstrated -- Incomplete bypassing -- to achieve similar clustering. Incomplete bypass based clustering is similar to traditional clustering in that it creates groups of functional units where intra-group communication occurs within a single cycle over fast bypass wires and inter-group communication takes longer, more than one cycle. One key difference is that in traditional clustered microarchitectures, inter-cluster communication takes place over the global buses whereas incomplete bypass designs achieve inter-group communication via the register file. It is demonstrated that incomplete bypass based clustered micro-architecture achieves higher performance (10% speedup) and better energy efficiency than traditional clustered microarchitectures.

Reputation-based Resilient Data Aggregation in Sensor Network

Carlos R. Perez — Thu, 06 Dec 2007 07:51:12 PST

Data aggregation, or the fusing of many sensor measurements into a single summary, has been proposed as an important primitive in wireless sensor networks. But data aggregation is vulnerable to security attacks and natural failures where a few nodes can drastically alter the result of the aggregation by reporting erroneous data. In this thesis we present RDAS, a robust data aggregation protocol that uses a reputation-based approach to identify and isolate malicious nodes in a sensor network. RDAS is based on a hierarchical clustering arrangement of nodes, where a cluster head analyzes data from the cluster nodes to determine the location of an event. It uses the redundancy of the data along with certain assumptions about the sensors and environment to determine what data should have been reported by each node. Nodes form part of a reputation system, where they share information about other node’s performance in reporting accurate data and use the reputation ratings to mitigate the effect of malicious nodes in the data aggregation. Our system is able to perform accurate data aggregation in the presence of individually malicious and colluding nodes. It also deals with attacks where nodes try to compromise the reputation system by reporting false accusation and false praise for other nodes. The system uses a separate metric called trust that captures the fidelity of a node in reporting on other nodes’ behavior. We present simulation results to show that the aggregation is more robust to security attacks than in the baseline case where all sensor nodes are treated equally.

Capacitance-voltage studies of atomic-layer-deposited MOS structrures on Gallium Arsenide and other III-V compound semiconductors

Tian Yang — Thu, 06 Dec 2007 07:48:37 PST

Si-based CMOS devices with traditional structure are approaching the fundamental physical limits. New device structures and materials must be explored to continue the trend of increasing electronic device speed and decreasing size at the same time. Recently, III-V compound semiconductors are considered as novel channel materials to replace Si due to their high electron mobilities. However, the main obstacle to implement III-V as novel channel materials for CMOS application is the lack of high-quality, thermodynamically stable insulators. Thus, systematic studies were carried out on atomic-layer-deposited (ALD) high-k dielectrics on GaAs and other III-V semiconductors. The experiment results show that interface quality strongly depends on surface pre-treatments, oxide materials and formation, post deposition annealing conditions, and also substrate semiconductors. A systematic interface study was also performed on ALD hafnium-aluminium-oxide laminate gate dielectric on GaAs. The results show that the hafnium-aluminium-oxide laminate structure gate dielectric improves the GaAs MOS characteristics such as dielectric constant, breakdown voltage and frequency dispersion. Due to small electron effective mass of GaAs, quantum capacitance effect is detrimental to the device performance. Indium tin oxide (ITO)gate was used to study quantum capacitance effect on GaAs. By using this transparent gate, strong inversion C-V is observed on GaAs MOS devices with ultra-thin dielectric layer. Capacitance value lowering was experimentally observed at electron accumulation.

Electro-Thermal Transport in Nanotube Based Composites for Macroelectronic Applications

Satish Kumar — Wed, 24 Oct 2007 12:31:44 PDT

Dispersions of particles of different shapes and sizes in fluids or solids modify the transport properties of the underlying matrix. A remarkable enhancement in the electrical, thermal and other transport properties of the matrix due to the long aspect ratio dispersions like nanotube/nanowires has been observed my many research groups. This has motivated tremendous research to explore these composites for various macro-electronic and micro-electronic applications in the last decade. Carbon nanotubes (CNTs) network based thin-film transistors (TFTs) promise improved performance for flexible plastic electronics with potential applications in displays, e-paper, e-clothing, bio-chemical sensing, conformal radar, and others. A detailed theoretical and numerical framework is required to understand the functioning of the CNT network transistors and to interpret different experimental observations.

In the present work we develop a computational model based on the classical transport equations to analyze the electro-thermal transport in isotropic 2D nanotube-net (Nanonet) based TFTs. We represent the Nanonet as a simple, two-dimensional, interpenetrating percolating network of metallic and semiconducting nanosticks. The methodology couples the electrical and thermal transport in an efficient and self-consistent manner. We show the effect of electro-thermal coupling on device performance and explore temperature rise as a function of different parameters like channel length (LC), network density, tube-to-substrate thermal conductance (BiS), and tube-to-substrate thermal conductivity ratio.

We also analyze the electrical characteristics of CNTs based organic thin-film-transistors. This technique relies on “doping” the organic host with metallic carbon nanotubes to increase the transconductance (equivalently, reduce effective channel length, Leff). Our analysis reproduces experimental characteristics and explains many trends not understood through the experimental observations. We show that Leff scales as a power-law of CNT-doping density and illustrate the importance of an active subpercolating network of semiconducting-CNTs in an organic host. To explore the viability and potential of this technology, we establish the upper limit of transistor-count for an IC based on this technology as a function of density, on-current and circuit-failure probability.

This work should be important both as a generalization of classical percolation theory to heterogeneous multi-component percolation as well as theory development and optimization of Nanonet transistors for flexible electronics.