Home // CENICS 2019, The Twelfth International Conference on Advances in Circuits, Electronics and Micro-electronics // View article

A Convolutional Neural Network Accelerator for Power-Efficient Real-Time Vision Processing

Authors:
Junghee Lee
Chrysostomos Nicopoulos

Keywords: Convolutional Neural Network; Hardware Accelerator; Scheduling

Abstract:
Deep Convolutional Neural Networks (CNN) constitute a promising framework for many applications. Such networks are often employed for vision processing algorithms, because CNNs offer better accuracy than traditional signal processing algorithms. However, it is challenging to apply high-accuracy deep CNNs for real-time vision processing, because they require high computational power and large data movement. Since general-purpose processors do not efficiently support CNNs, various hardware accelerators have been proposed. While it is required to support all the layers of the CNN for real-time vision processing, the large amount of weights (more than 100s of MB) limit the speedup of hardware acceleration, because the performance is largely bounded by memory access times. Recent CNN architectures, such as SqueezeNet and GoogLeNet, address this problem by employing narrow layers. However, their irregular architecture necessitates a re-design of hardware accelerators. In this paper, we propose a novel hardware accelerator for advanced CNNs aimed at realizing real-time vision processing with high accuracy.

Pages: 25 to 30

Copyright: Copyright (c) IARIA, 2019

Publication date: October 27, 2019

Published in: conference

ISSN: 2308-426X

ISBN: 978-1-61208-748-1

Location: Nice, France

Dates: from October 27, 2019 to October 31, 2019