Year: 2019
Type: dissertation
Abstract: Computational Imaging (CI) is an essential front-line building block for a wide range of applications including computer vision, advanced driver-assistance systems, medical imaging, photography and robotics. Such applications process large amounts of sensory data and need to do so either in real-time or within user-interactive setup constraints. In addition, these applications are typically run on cost- and energy-constrained devices with limited compute and memory resources. CI algorithms perform sophisticated processing of the noisy sensor output to compensate for the imperfections of the image acquisition hardware, to overcome challenging imaging conditions and to reveal new features of the captured image. This dissertation analyzes the performance of recent CI implementations on high-end commodity processors showing that practical deployment is beyond the reach of such architectures. Accordingly, the dissertation identifies unique characteristics of the data processed by these algorithms and exploits them to propose efficient data-aware CI accelerators. Specifically, this dissertation accelerates two CI implementation approaches: 1.Analytical solutions represented by Block-Matching and 3D filtering (BM3D), a state-of-the-art algorithm that is used in a wide set of core CI tasks including image denoising, sharpening, deblurring and super-resolution. We propose IDEAL, an image denoising/sharpening hardware accelerator. IDEAL incorporates Matches Reuse (MR), a novel software-hardware co-optimization that exploits typical image content to significantly reduce the computations needed without loss in output quality. IDEAL is orders of magnitude faster and more energy efficient compared to high-end CPUs and GPUs. 2.Machine learning-based implementations represented by a set of recent deep neural networks (DNNs) performing several CI tasks. We characterize strong spatiotemporal correlation between the adjacent values processed by these DNNs. Accordingly, we introduce Differential Convolution, a convolution processing technique that operates on the deltas of the values rather than on the absolute values. We propose Diffy, a novel DNN accelerator that incorporates Differential Convolution translating the characterized correlation into improved performance, reduced on- and off-chip storage and communication, and ultimately, improved energy efficiency. The proposed hardware accelerators enable practical deployment and achieve the performance needed for user-interactive, real-time and continuous vision applications. The designs are shown to outperform state-of-the-art accelerators and commodity architectures.
Source: TSpace (University of Toronto)
Author: Mostafa Mahmoud Hassan Mahmoud
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3090900754