Search Constraints
Filtering by:
Campus
Fullerton
Remove constraint Campus: Fullerton
Department
Department of Computer Science
Remove constraint Department: Department of Computer Science
« Previous |
1 - 20 of 30
|
Next »
Number of results to display per page
Search Results
- Creator:
- Yu Bai
- Description:
- During recent years, researchers throughout academia and industry have been advancing the theory, designing, and applications of mobile service computing through the Internet of Things (IoT). Research interest in mobile service computing stems from its performance, security, reliability, and power consumption. Hence, ultra-low power integrated circuits are essential for mobile service computing that can offer the advantages of low power for computational tasks in the IoT that is driven by the restricting constraints of power consumption and autonomy in both computation and idle phases. To attain the benefits of ultra-low power circuits, the energy-consuming and computational intense demands are imposed by the underlying processing and memory devices on which the conventional ultra-low power integrated circuit can benefit substantially from innovative hardware designs. Logic-in-Memory (LIM) architectures are considered as the potential approaches to attain goals within area and energy constraints starting with the lowest layers of the hardware stack. In this paper, we propose and implement the LIM asynchronous computing paradigm for energy-efficient mobile service computing. The results indicate that the proposed design achieves 38% leakage reduction and 30% accuracy improvement compared to the state-of-the-art non-volatile asynchronous circuits. At the system level, we compare our designs with various commercial microprocessors. The experimental results show that the asynchronous processors attain a four-fold throughput increase relative to their synchronous counterparts under these operating constraints. Therefore, the proposed design offers an approach toward tangible benefits of the battery-constrained embedded mobile service computing.
- Resource Type:
- Article
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Vo, Holly H.
- Description:
- Diabetic retinopathy (DR) is a common eye disease that could lead to irreversible vision loss but hard to be noticed by carriers in early stages. On the path of recognizing DR stages by multi-scale color uniform local binary pattern in retinopathy images, this work explores two main point. The first point is investigating the role of feature dimensionality reduction in the process of extracting discriminatory features for effective classification. The second point is exploring the discriminatory information carried in different color spaces for fundus images. Experiments are conducted on a large scale dataset of 35,126 training images and 53,576 testing images that have been taken by different devices with high variance in dimension, quality and luminance. The proposed multi-level feature dimensionality reduction (FDR) methodology is applied in three scopes in the feature hierarchy of the fusion of five color spaces: RGB, L*a*b*, HSI, I1I2I3, and rgb. The novel combination of the proposed multi-level FDR method and color fusion achieves 75.2% accuracy by one-to-one SVM classifier.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Rathod, Priyank Jayantilal
- Description:
- Recently there is a growing trend towards the emotional information to be used in the human-computer interaction. In order to get the correct interaction from the computer, a human has to provide the appropriate inputs. Inputs are recorded by various bio-medical sensors and processed for generating the appropriate data. For examples, Skin conductance is used to measure the skin potential between two nodes in the human body whereas heart rate as the name suggests is used to measure the variation in heart rate. Biomedical signals are generated using a neural network in a human body and used to estimate the emotions. As these signals are generated using the neural network, it is not possible to control it artificially. Due to this reason, it is the reliable source of the estimation of such information. Heart Rate Variability (HRV) and Galvanic Skin Response/Conductance (GSR) is changed when any type of emotion is induced in the human body. This change in signals represents certain characteristics which are used to estimate the emotions. Mainly there are two types of emotions which are negative emotions and positive emotions. Positive emotion includes happiness and normal behavior whereas negative emotion includes fear, anger and sadness. The current study investigated five different types of emotions. This research consists of measurement procedures for the emotion detection using biomedical signals.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Shinde, Nikhil Hari
- Description:
- Augmented Reality (AR) is a well-known and highly evolving technology in the past few years with a mixture of real and virtual environment. The AR technology has attracted lots of attention in different field such as engineering, medical, and digital. Computer generated animations and visualizations are presented to the user in the real world with an object detection, tracking, and by augmenting it. AR has been intensively researched over several decades with the purpose of being a great advantage in different fields but has yet to reach a specific purpose to a common audience. As a consideration AR is more over suitable to visualize the context related information but in a broader way could be utilized in the form of a control system by pointing a specific registered object or graphics with a real time application. This thesis presents efforts in bringing Augmented Reality to the biomedical world as a driving aid for electric wheelchairs. Implementing a technology on newly evolved devices like AR glasses and to activate it for a purpose of controlling any hardware system poses a large number of challenges that differ from the requirements of research in traditional purposes. These include: limited device interface resources with little or no possibility to upgrade or add hardware and limited capabilities and functionalities of the AR technology in the area of controlling the hardware systems. The research presented in this thesis addresses these challenges and makes contributions in the following area: Augmented Reality based driving aid for electric wheelchairs.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Al-barazanchi, Hussein Ali Kadhim
- Description:
- Plankton are a diverse group of organisms that live in large bodies of water. They are a primary source of food for fishes and other larger aquatic organisms. The distribution of plankton plays a significant role in the marine ecosystem. Therefore, the study of plankton distribution is an important tool used for assessing the changes in the marine ecosystem. The study of plankton distribution is based mostly on the classification of plankton images taken by underwater imaging systems. The images used in this study come from the SIPPER system. The challenges with SIPPER’s plankton image dataset are the low resolution of images, the high degree of similarities between different classes, high variability within the same class, partial occlusion, and noise. Also, traditional computer vision techniques require tedious work to find suitable features to represent plankton. Having a robust automated system for classification of plankton images will play a significant role in advancing marine biology research. To overcome those issues, we propose the use of Convolutional Neural Networks (CNN) and Hybrid-CNN for this task. Results of the experiments on SIPPER dataset show improvement in classification accuracy in comparison with other states of the art approaches. Another major advantage of this approach is the scalability for classification of new classes without the need for features engineering.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Dasamoni, Thanga Nirmaia
- Description:
- Many model-based approaches have been suggested to improve and accelerate the development process of complex distributed embedded systems (DES) from design to deployment. In this approach, abstract models of the system are created and then transformed systematically into concrete implementations. Modeling focuses on commands and information being distributed reliably throughout the system with minimized latency to maintain acceptable QoS. In this research, an irrigation system is developed. Model-based development was used to model the architecture and communication of distributed nodes. However, the complexity of the Bluetooth Low Energy (BLE) communications used by the irrigation system requires a model with very high sophistication in the bottom-up modeling approach. A distributed irrigation system is implemented as a proof of concept using an experimental Bluetooth Low Energy based network topology and protocol, and the data was observed to verify the correct operations.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Rahimi, Ahdel
- Description:
- Identifying the type of modulation scheme present on a signal is an integral part of any non-coherent receiver, especially in defense. Prior to signal processing, modulation schemes present on a received signal must be accurately identified. Finding a fast and accurate method for classifying the signal type can be very difficult. The use of machine learning algorithms has been explored as a solution for signal classification; however, many machine learning algorithms and their proposed applications are inadequate for real-time processing due to the complexity and overhead associated with the algorithms chosen. Here a supervised machine learning approach using a k-Nearest Neighbor (kNN) algorithm is efficiently applied to achieve very accurate modulation classification using only three features extracted from the signal. The use of a simpler machine learning algorithm such as kNN and a reduced feature set for classification allows for various optimization techniques to be applied in hardware for classification of a received signal. With the presented classification method, a hardware implementation can achieve the exact or better results than the simulated system by loading all of the classification data into content-addressable memory (CAM) based look-up table, and directly mapping a possible outcome to a particular group for classification. Ternary content-addressable memories (TCAMs) may also be used to help optimize this approach further through simplified search queries.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Sajit, Ahmed Sattar
- Description:
- The 21st century is witnessing a tremendous demand for transistors. Life amenities have incorporated the transistor in every aspect of daily life, ranging from toys to rocket science. Day by day, scaling down the transistor is becoming an imperious necessity. However, it is not a straightforward process; instead, it faces overwhelming challenges. Due to these scaling changes, new technologies, such as FinFETs for example, have emerged as alternatives to the conventional bulk-CMOS technology. FinFET has more control over the channel, therefore, leakage current is reduced. FinFET could bridge the gap between silicon devices and non-silicon devices. The semiconductor industry is now incorporating FinFETs in systems and subsystems. For example, Intel has been using them in their newest processors, delivering potential saving powers and increased speeds to memory circuits. Memory sub-systems are considered a vital component in the digital era. In memory, few rows are read or written at a time, while the most rows are static; hence, reducing leakage current increases the performance. However, as a transistor shrinks, it becomes more vulnerable to the effects from radioactive particle strikes. If a particle hits a node in a memory cell, the content might flip; consequently, leading to corrupting stored data. Critical fields, such as medical and aerospace, where there are no second chances and cannot even afford to operate at 99.99% accuracy, has induced me to find a rigid circuit in a radiated working environment. This research focuses on a wide spectrum of memories such as 6T SRAM, iii 8T SRAM, and DICE memory cells using FinFET technology and finding the best platform in terms of Read and Write delay, susceptibility level of SNM, RSNM, leakage current, energy consumption, and Single Event Upsets (SEUs). This research has shown that the SEU tolerance that 6T and 8T FinFET SRAMs provide may not be acceptable in medical and aerospace applications where there is a very high likelihood of SEUs. Consequently, FinFET DICE memory can be a good candidate due to its high ability to tolerate SEUs of different amplitudes and long periods for both read and hold operations.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Kuerban, Mutalifu
- Description:
- Software-defined networking (SDN) introduces centralized control logic, and it separates the data plane from the control plane. Hence, SDN makes it easy for network engineers to monitor traffic, diagnose threats, and insert or change security policies. Also, SDN makes the network more flexible and customized integration easier. However, it also creates security challenges which did not exist before, such as controller security. In this paper, we use a layered approach to describe SDN structures. In this thesis, the effect of Denial of Service (DoS) attacks on the SDN controller has been analyzed and a strategy to mitigate DoS attacks has been suggested. An analysis has been provided by emulating an attack on an SDN. The SDN uses a network created by Mininet and Floodlight as a controller.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Hosseini, Seyed Mohammad Mahdi
- Description:
- Considerably deep convolutional neural networks have contributed greatly to recent advances in image recognition. the emergence of architecture like GoogLeNet with Inception Module have allowed neural networks to gain high performance with low computational cost. It is possible to have substantial gains in classification tasks but this directly equates to increased parameter count and computational efficiency trade-offs which Inception architecture addresses. in addition, architectures such as VGG-19 with their considerable depth and small 3 × 3 convolution filters allow us to achieve very high accuracies. the task of car model classification is a challenging, fine-grained task and cars can have broad difference between different makes but at the same time subtle differences in comparison to cars within the same make. in addition, the pose of the car image samples in the dataset as well as the setting the images are captured with can make this task even more challenging. Architectures that are deep and have many layers are best suited for learning such hierarchies. to this end, in this paper, we propose a new model that relies on fusion of VGG and GoogLeNet layers with trained weights on ImageNet dataset allowing us to classify car images at a high rate of accuracy.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Mahmud, Ali
- Description:
- I attempt to use a model of combining features of convolutional neural networks to classify cars from the Comprehensive Cars Data set. I try to improve on the single model approach by fusing features from multiple convolutional neural networks. the networks in question that I tried to fuse in this way are: ResNet, GoogLeNet, and CNDS. Initial training procedures showed promise with expected high accuracy, however, I was unable to successfully fuse the networks. I was not able to classify images extracted features with a support vector machine or random forest.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Ligon, Jason Marco K.
- Description:
- Due to the increasing amount of interest in autonomous systems, such as self-driving cars and self-navigating drones, object detection has become an increasingly important and challenging task. 3D point clouds are commonly acquired from LiDAR (Light Detection and Ranging) and 3D laser scanners such as the Microsoft Kinect. This thesis proposes a step-by-step methodology of 3D point cloud processing for object detection. the method can be broken down into four major steps: downsampling, surface normal calculation, descriptor, and feature matching. Downsampling involves the reduction of points in a point cloud. Calculating the surface normal allows for the determination of the orientation of each point thus giving us information about the shapes of the object. Descriptors provide feature points and are used as a means to “describe” an object to a computer. Feature matching involves the use of pre-calculated feature points and determines whether the object the computer is looking for is in the scene. Since the use of point clouds is relatively new for object detection, this allowed for the create a simple framework to handle an enormous amount data. Though our proposed pipeline relies on the use of spin images for the descriptor stage, it lends itself to have easily interchangeable algorithms at every step. Finally, the proposed methodology had scalability in mind. Therefore, the framework performs within a reasonable amount of time which is achieved by reducing the number of calculations at each step.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Dasamoni, Thanga Nirmaia
- Description:
- Many model-based approaches have been suggested to improve and accelerate the development process of complex distributed embedded systems (DES) from design to deployment. In this approach, abstract models of the system are created and then transformed systematically into concrete implementations. Modeling focuses on commands and information being distributed reliably throughout the system with minimized latency to maintain acceptable QoS. In this research, an irrigation system is developed. Model-based development was used to model the architecture and communication of distributed nodes. However, the complexity of the Bluetooth Low Energy (BLE) communications used by the irrigation system requires a model with very high sophistication in the bottom-up modeling approach. A distributed irrigation system is implemented as a proof of concept using an experimental Bluetooth Low Energy based network topology and protocol, and the data was observed to verify the correct operations.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Yenter, Alec
- Description:
- There has been tremendous victory and improvement in deep learning with the use of neural networks. Areas of image and video have benefited from this advancement, but there is room for further development in the field of sentimental analysis. Specifically, there is an abundance of text reviews of movies that require insightful classification of sentiment. This thesis first reviews machine learning literature to understand the current performance on a movie review dataset acquired from IMDb.com. Second, a combined-kernel convolutional-based Long Short-Term Memory network is proposed to perform deep learning on the challenging data. the paper also analyzes multiple approaches and variations of the model. the proposed network can achieve the highest known accuracy on the IMDb review sentiment dataset at 89.5%. the network’s success can be extended to further other fields.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Patnam, Venkata Sindhoor P.
- Description:
- Children with autism often experience sudden meltdowns which not only makes the moment tough for the caretakers/parents but also make the children hurt themselves physically. Studies have discovered that children with autistic spectrum disorder exhibit certain actions through which meltdowns in them can be anticipated. the objective of this project is to build a system that can recognize such kind of actions using deep learning techniques thereby, notifying the caretakers/parents so that they can get the situation under control in lesser time. Using deep learning RCNNs, the system can be trained faster yet reliable because unlike all the machine learning algorithms, deep learning algorithms are more efficient and have more scope for future. a classifier is trained on images that are gathered from videos and reliable internet sources with most predictive gestures, through which the meltdowns can be detected more precisely. a model that validated the accuracy by ~93% which is accompanied by a loss/train classifier with a minimal 0.4% loss is trained. Functional testing was done through feeding the deep neural network with chosen actions performed by five individuals that resulted in an accuracy of ~92% in all cases, which can assure the real-time usage of the system.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Mistry, Krupal Sureshbhai
- Description:
- A Brain-Computer Interface (BCI) provides an interaction between a brain and a device that enables the electroencephalography (EEG) signals from the scalp of the person to control an external device, such as operations of an electric wheelchair, electronic devices, etc. This study provides a Steady State Visually Evoked Potential (SSVEP) based BCI system to control the operation of an electric wheelchair by using the EEG signals obtained from the brain. the main objective is to process the EEG signals obtained from an externally driven stimulus and trigger the control signals to operate the electric wheelchair. by providing attention to the external visual stimulus, corresponding EEG signals are elicited from the visual cortex region of the brain. the obtained EEG signals are classified at different frequencies using signal processing algorithms and given as an input to the BCI system, which controls the operation of an external device (electric wheelchair). the proposed application provides a platform for the individuals suffer from Neuromuscular Degenerative Diseases (NMDs) such as Amyotrophic Lateral Sclerosis (ALS), Locked-In Syndrome (LIS), etc., and help them to lead an independent life. Four trials have been performed to measure the accuracy and reliability of the system. Also, the proposed paradigm is compared with the Audio Steady State Response (ASSR) approach.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Geethakumari Anil, Divya
- Description:
- A Brain-Computer Interface (BCI) provides an interaction between a brain and a device that enables the electroencephalography (EEG) signals from the scalp of the person to control an external device, such as operations of an electric wheelchair, electronic devices, etc. This study provides a Steady State Visually Evoked Potential (SSVEP) based BCI system to control the operation of an electric wheelchair by using the EEG signals obtained from the brain. the main objective is to process the EEG signals obtained from an externally driven stimulus and trigger the control signals to operate the electric wheelchair. by providing attention to the external visual stimulus, corresponding EEG signals are elicited from the visual cortex region of the brain. the obtained EEG signals are classified at different frequencies using signal processing algorithms and given as an input to the BCI system, which controls the operation of an external device (electric wheelchair). the proposed application provides a platform for the individuals suffer from Neuromuscular Degenerative Diseases (NMDs) such as Amyotrophic Lateral Sclerosis (ALS), Locked-In Syndrome (LIS), etc., and help them to lead an independent life. Four trials have been performed to measure the accuracy and reliability of the system. Also, the proposed paradigm is compared with the Audio Steady State Response (ASSR) approach.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- O'bard, Bryce
- Description:
- Supervised machine learning has become a powerful method of creating predictions and classifying labeled data sets. the ability to analyze and classify data is a powerful tool used in applications spanning almost every industry. the most popular classification algorithms include support vector machines (SVM), K-nearest neighbors (KNN) and decision trees. These algorithms separate labeled datasets based on features extracted from the inputs. in the assistive technology field, the use of eye tracking technology provides patients with quadriplegia, ALS, or other neurodegenerative diseases with the ability to control speech devices using their eyes. to provide a low-cost alternative to the existing costly devices, an electrooculography (EOG) Bluetooth controller was developed for patients to control an iOS device hands-free. the classification of looking up, down, left, and right was implemented using threshold limits and signal processing logic without machine learning. This methodology fails, however, to filter out unintentional blinks and record intentional blinks. a dataset consisting of these gestures was collected over several trials and classified using an SVM, KNN, and decision tree’s algorithm with a moving window buffer size suitable for an embedded device. the trained algorithms are converted to C code and uploaded to an ATmega328p AVR microcontroller. Using a decision tree implementation, the intentional blink signal classification is successfully predicted with an accuracy of 97.33% accuracy and filters out unintentional blinks with 100% accuracy on the embedded device.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Kim, Daniel E.
- Description:
- In this thesis, the concept of neural networks, a form of machine learning, is applied to network intrusion detection to compare the performance of shallow and deep neural networks. Neural networks provide a robust method of machine learning that can identify patterns and classify observations or objects. Several structures of shallow and deep networks are tested with varying hyperparameters to create a good understanding of the performance and capability of each type of network. Using currently available labeled datasets, our experiments and evaluations show that shallow neural networks are able to more successfully identify malicious network traffic than the more complex deep neural networks, with the former achieving a peak average performance of 98.50% detection rate and the latter only reaching an average high of 48.30% detection rate.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science
- Creator:
- Ly, Phillip
- Description:
- The goal of this thesis is to create mobile applications that can leverage the power of deep learning to detect malignant melanoma in the early phase and save lives. Thus, it is imperative to extend the reach of such essential diagnostic care worldwide. in this thesis, we will first present three deep learning methodologies that entail constructions of convolutional neural networks in conjunction with the uses of modern transfer learning and regularization techniques. the proposed deep learning methodologies leverage a dynamic dataset to optimize performance of a skin cancer classification mobile application called ChekSkin. Dynamic datasets refer to the expansion of datasets from influx of new data. Furthermore, the proposed deep learning methodologies generate mobile compatible models by rendering and training 80,192 high quality images. We performed rigorous experiments to attain the following top-1 accuracies: 81% (overall accuracy on the test dataset) using advanced transfer learning and data augmentation techniques via TensorFlow, 85.7% by training a batch-normalized CNN from scratch, and 88.35% with the uses of potent feature extraction and data augmentation methods via Keras. Additionally, the ChekSkin app is tested in real-world situations in which there are drastic variations in lighting conditions and image quality. We have considered tests in both experimental and real-world settings as important metrics for life-saving mobile applications.
- Resource Type:
- Masters Thesis
- Campus Tesim:
- Fullerton
- Department:
- Department of Computer Science