There is a tremendous demand for increasingly efficient ways of both capturing and processing high-dimensional datasets of large size. When capturing such datasets, a promising recent trend has developed based on the recognition that, many high-dimensional datasets have low-dimensional structures. For example, the notion of sparsity is a requisite in the compressed sensing (CS) field, which allows for accurate signal reconstruction from sub-Nyquist sampled measurements given certain conditions. When processing such datasets, the recently developed deep learning is a powerful tool, able to extract high-level and complex abstractions from massive amounts of data. CS has a wide range of applications that include imaging, radar and many more. Much effort has been put on developing more accurate and efficient reconstruction algorithms. In this thesis, first, we are interested in how to incorporate the side information into CS reconstruction when there is an initial estimation of the sparse signal available from other sources. Rigorous theoretical analysis was proposed for the first time in this field. Sufficient number of measurements is required for accurate CS reconstruction. We may have to wait for a long time to do the reconstruction until we receive enough measurements, which could incur undesired delays. Moreover, state-of-the-art CS reconstruction algorithms are still inefficient for signals of large size, e.g., images. Inspired by the multi-resolution or scalable reconstruction in multimedia transmission, such as JPEG 2000 and H.264/SVC, in the second part of this thesis, we analyzed scalable CS reconstruction problem and proposed to reconstruct a low-resolution signal if the number of measurements is too small. Deep learning or deep neural networks (DNNs) has evolved into the state-of-the-art technique for many artificial intelligence tasks including computer vision, speech recognition and natural language processing. However, DNNs generally involve many layers with millions of parameters, making them difficult to be deployed and updated on devices with limited resources such as mobile phones and other smart embedded systems. Moreover, if the DNN needs to be updated, usually via wireless communications, downloading the large amount of network parameters will cause excessive delay. In the final part of this thesis, we propose a scalable representation of the network parameters, so that different applications can select the most suitable bit rate of the network based on their own storage constraints.
In recent years, Optical Coherence Tomography (OCT) has become one of the dominant imaging technologies for ophthalmic diagnostics and vision research. The fast and high-resolution cross-sectional data that OCT provides has brought a new possibility in the role of intra-operative imaging. However, existing commercial OCT systems lack the automated real-time functionality for providing immediate feedback of changes in anatomical configuration as the result of surgical actions. The predominant reason for lacking such functionality is because high complexity algorithms are hard to implement in real-time imaging due to their computationally expensive nature. In this thesis, we will present a Graphics Processing Unit (GPU) accelerated retinal layer segmentation for real-time intra-operative imaging applications. Modern GPUs has emerged as a strong tool for mass computation in scientific researches. The computational power of the GPU outpaces Central Processing Unit (CPU) significantly when the processing task is parallelizable. Image segmentation is a computationally expensive algorithm and traditionally implemented in sequential instructions. An example of a parallelizable segmentation algorithm is Push-Relabel (PR) Graph-Cut(GC), which can be implemented using GPU. The GPU Retinal Segmentation (GRS) presented in this thesis is built upon such an algorithm. To ensure the run time of the GRS meets the real-time requirement for its application, multiple GPUs are used to accelerate the segmentation processing further in parallel. As a result of using GRS, we were able to achieve the visualization of the retinal thickness measurement and the enhancement of retinal vasculature networks in real-time.
Seismocardiography is the non-invasive measurement of the heart vibration by placing an accelerometer on the human chest. Due to its non-invasive nature, the seismocardiogram signal could be embedded inside portable devices for the purpose of health monitoring and remote diagnosis. With the combination of the electrocardiogram (ECG) signal, cardiac time intervals (CTI) could be extracted. CTIs are timing intervals that are associated with specific events of the cardiac cycle. The research community has explored the potential of CTI in the diagnosis of chronic myocardial disease, ischemic and coronary artery disease, arterial hypertension, cardiac resynchronization therapy, and implantable cardioverter de-fibrillator. For the extracted CTIs to be useful in a medical device, the seismocardiogram signal (SCG) has to be automatically delineated. Upon the automatic delineation of CTIs, the timing parameters could be either combined with other physiological signals to create new indices that have unique physiological interpretations or to be used as a complimentary technology. Hence, The present dissertation has three main objectives: (1) automatic SCG delineation algorithm, (2) application of cardiac time intervals (extracted from SCG) for generating aunique index for early stage hemorrhage detection, and (3) complementary technology for optimization of the diastolic timed vibration therapy. For the first objective, the proposed delineation algorithm had the capability to correctly estimate the CTIs while discarding low-quality cardiac cycles, which are the ones thatdon’t have identifiable fiducial points. For the second objective, the combination of the electrocardiogram, seismocardiogram, and photoplethysmogram signals was used to design a hemorrhage progression index, which ultimately was applied for early stage detection. For the last objective, the extracted CTIs were applied to the “diastolic timed vibration”, which is a potential therapy for patients with acute ischemia during the pre-hospitalization phase. A calibration methodology was proposed for diastole detection in real-time.
Sound Field Reproduction (SFR) is for creating a desired sound field from a primary source by using multiple loudspeakers. Its research calls for numerical experiments by simulation. SFR performance can be improved by optimizing the static Degrees of Freedom, i.e., the locations and patterns of loudspeakers. The approach is to decrease the sound field reproduction error without increasing the operational complexity, and this requires that the possible locations and frequencies of the primary source are known a priori. To optimize the loudspeaker locations, two placement methods are developed. In the first method, an idealized Acoustic Transfer Function (ATF) matrix that minimizes the reproduction error, but which may not be realizable, is derived for a fixed number of uniformly placed, omnidirectional loudspeakers. The loudspeakers are then re-positioned within their aperture so that their realizable ATF matrix best approximates the idealized ATF matrix. In the second method, a new algorithm is called – Constrained Matching Pursuit (CMP), which optimizes loudspeaker location while constraining the total loudspeaker power to avoid acoustic hotspots. CMP is also used to jointly optimize the radiation patterns and locations of the loudspeakers. These methods optimize for a single frequency, and a method is presented which extends the case to multiple frequencies for the primary source. The multi-frequency method is deployed in the audio layer of an immersive communications system. An existing model for the Head Related Transfer Function (receiving pattern) is adapted for the computation of the loudspeaker excitation functions, called the dynamic Degrees of Freedom. Subjective and objective tests are applied and concur that the quality of speech of the SFR in a reverberant room is significantly improved compared to a system with the same number of loudspeakers that are uniformly spaced and omni-directional, and which have the same total power constraint and computation complexity.
This work demonstrates a novel means of manufacturing nano-optical devices, functioning according to the principle of structural colouration, by inkjet printing silver nanoparticle ink on nanostructured surfaces. The structural colouration is created by a surface containing micro- or nano- features interacting with light. In this study, we use a polymer substrate patterned with different types of nanostructure arrays as structural pixels that give red, green, and blue primary colours. We utilize nanoimprint lithography to replicate the nanostructured substrate from a prefabricated stamp. These days, inkjet printing has become a scalable micropatterning technology due to its precise, flexible control and cost-effective additive process. In this current work, inkjet printing technology is employed to selectively activate pixels by printing silver nanoparticle ink on the nanosubstrate surface to gain colour mixing. In the experiments performed, a nanostructured substrate patterned with diffractive nanostructure arrays was implemented to print full-colour images. The effect of surface wettability, different concentrations of silver nanoparticle ink, drop and line spacing, and polymer nanostructures on the optical properties of the subpixels of dried silver printed dots, are investigated to achieve high printing resolution. This method is a key to achieving full-colour, scalable, high-throughput, flexible and cost-effective printing of structural colour images. The printed pictures demonstrate unique optically variable effects that distinguish from security products using pigments or printing inks in their manufacturing processes. Therefore, this technique is an ideal candidate for security and authentication applications that require customizable anti-counterfeiting features.
Work presented in this thesis demonstrates methods of combining a newly developed magnetic composite polymer (M-CP) with other commonly used polymer microfluidics materials for the creation of complex all-polymer microfluidic systems. To achieve fully integrated microfluidic systems, new fabrication techniques for integration of M-CP structures are developed. Employing the new M-CP material and the novel fabrication techniques, three types of actuators are developed: cilia, flap, and hybrid M-CP/PDMS actuator. All three actuators employ compatible materials, fabrication techniques, and actuation mechanisms. The performance of each of these actuators is characterized for different applications: cilia-based mixers, flap-based valves, and hybrid M-CP/PDMS actuators for applying extracellular stimulation on cell monolayers. The actuators in each of these applications are driven via relatively small external magnetic fields. The M-CP used in these novel actuators is composed of rare-earth magnetic micro-particles (5–10 micrometer) that are embedded in polydimethylsiloxane. The M-CP is patterned into large force, large stroke actuators. The polymer matrix without magnetic particles is employed as the substrate material for passive parts, facilitating integration of the magnetic and non-magnetic materials. The compatible fabrication techniques include a modified soft-lithography technique for hybrid M-CP/PDMS actuators, screen printing via shadow masks for micro-patterning of thin layers of M-CP, and a novel fabrication technique using poly(ethylene glycol) (PEG) as a sacrificial material for the fabrication of ultra-high aspect-ratio and highly flexible M-CP cilia. Microfluidic devices using these actuators show improved performances in their respective fields when compared with existing designs. Microfluidic mixers with 8 cilia show a reduction in mixing time of up to 63 times over diffusion. Flap-based valve arrays effectively switch flows between two microfluidic channels using an array of two valves, and effectively perform as on-off switches for flow control. A valve with a 2.3 mm flap thickness, actuated under an 80 mT magnetic field, is capable of blocking liquid flow at a flow rate of 1 mL/min for pressures up to 9.65 kPa. Microfluidic platforms for stretching/compressing biological cells based on the hybrid M-CP/PDMS actuators achieve large and bi-directional surface deflections. Actuation can be applied cyclically, under both flow and no-flow conditions.
The main organ of the human central nervous system, the human brain, is one of the most complex organs in the human body. The causes of many brain diseases and disorders, such as Alzheimer’s disease, and their ideal treatments are still not fully understood with the current medical technology. With medical imaging techniques such as magnetic resonance imaging (MRI), magnetoencephalography (MEG), and electroencephalography (EEG), the data obtained from these techniques can be used to study and examine brain diseases and disorders. This project focuses on utilizing a surface registration method on multiple brain surfaces to obtain various geometric transformations for brain studies, and the implementation of the analysis pipeline on a high performance computing (HPC) environment. Due to the infeasibility on runtime for performing surface registration between one template brain surface and multiple target brain surfaces, an approach to perform sub-surface extraction on each brain surface and computation on a HPC environment has been employed. This has allowed a significant reduction in runtime and has also allowed the results to be obtained within reasonable time.
This paper discusses ion wind based on corona discharge. Ion wind generated by corona discharge voltage and an electric field can be used as a cooling device to replace a ventilation fan. Like some prior works done by others, the device is mainly composed of thin wires and blunt brass bars. The wind velocity can reach a value of 1 m/s under 3000 volts. In this report, the relationships between velocity and many parameters have been determined.
Long Term Evolution (LTE) is a standard for wireless communication developed by the 3rd Generation Partnership Project (3GPP) with an aim to fulfill the requirements defined for the fourth generation (4G) wireless networks. With more than hundred service providers across the globe and around one billion subscribers predicted by the year 2016, LTE is set to become the first true global standard. With the high data rates supported by LTE, improvements like Content Distribution Network (CDN) and increase in router switching speeds, popularity of video streaming services over the mobile networks is set to touch a new high. This popularity provides opportunity to the network operators to increase their revenues, but it also challenges them to provide video streams with minimum desirable quality to their customers. This has led to the emergence of two popular single layer video coding standards namely, H.264/AVC and more recently H.265/High Efficiency Video Coding (HEVC). With LTE being projected as a candidate to fuel the future 4G services, it is desirable to evaluate the performance of these two video coding standards over the LTE networks. The first part of this project tries to evaluate the video quality offered by these two video coding standards over LTE network by studying the impact of delay, distance and number of users in the LTE cell. In the second part of this project we implement a frame dropping mechanism which drops low priority frames of the video encoded with hierarchical B-frame structure when the channel conditions are not ideal, thus providing graceful degradation to the single layer videos. This mechanism tries to exploit the fact that in a video that is encoded using a hierarchical structure, the loss of a frame that belongs to the higher indexed temporal layer of a video has less negative impact on the video quality in comparison to the loss of a frame in the lower indexed temporal layer.
In today’s world, people are widely using technology to make their lives more comfortable and better. The development of semiconductors technology is making Integrated Circuits(IC) smaller and smaller in size, thus allowing IC designer to include more and more functionalities in their products. This development of technology has allowed a large diffusion of semiconductor devices in all aspects of human life, leading to the concept of “embedded” computation, described as the practice of including the small processor devices in all spaces of our world, from our houses, to our cars, to even “wearable electronics” that we carry around as we move. In particular, floating point computation (FP) is a feature of computers that, at the price of significant additional hardware complexity and sometimes at the price of result accuracy, provides a much larger range of usable numbers, thus significantly enhancing the flexibility and usability of our computation. The additional hardware complexity imposed by FP units imposed a relevant price in Silicon Area (making the IC more expensive) and especially in terms of power consumption. In turn, energy consumption is a very severe issue in semiconductor technologies: first, it causes unreliability of the IC technology. Secondly, IC energy consumption leads to greenhouse gas emission. Finally, many IC systems are battery operated and high consumption may jeopardize the system usability and/or user experience. One very significant category of embedded processors is that of embedded sensors. Embedded sensors produce relevant quantities of raw data that needs to be adequately classified in order to provide significant information, and Machine Learning is often applied as a strategy for sensor data classification.This MENG project aims at exploring design strategies for low-power FP computation. In the following, we will introduce the design of a hardware FPU unit whose sub-blocks can be programmed to change dynamically the computational speed with the change in the voltage. This enables the FPU to adapt their consumption to the requirement of the environment, offering high performance (and high consumption) whenever needed by the environment, but adapting to low power, low speed mode whenever intensive processing is not necessary.