Convolutional neural network acceleration with hardware/software co-design

Chen, Andrew Tzer-Yeu; Biglari-Abhari, Morteza; Wang, Kevin I‐Kai; Bouzerdoum, A.; Tivive, Fok Hing Chi

doi:10.1007/s10489-017-1007-z

Cited by 8 publications

(9 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…[167]), and neural networks (see e.g. [168]). Algorithms from those domains are cornerstones of various biometric systems; hence, substantial research effort has also been devoted to development of FPGA‐based processing of biometric data.…”

Section: Computational Workload Reduction Approachesmentioning

confidence: 99%

Computational workload in biometric identification systems: an overview

2019

View full text Add to dashboard Cite

Computational workload is one of the key challenges in biometric identification systems. The naïve retrieval method based on an exhaustive search becomes impractical with the growth of the number of the enrolled data subjects. Consequently, in recent years, many methods with the aim of reducing or optimising the computational workload, and thereby speeding-up the identification transactions, in biometric identification systems have been developed. In this article, a taxonomy for conceptual categorisation of such methods is presented, followed by a comprehensive survey of the relevant academic publications, including computational workload reduction and software/hardware-based acceleration. Lastly, the pertinent technical considerations and trade-offs of the surveyed methods are discussed, along with an industry perspective, and open issues/challenges in the field.

show abstract

Section: Computational Workload Reduction Approachesmentioning

confidence: 99%

Computational workload in biometric identification systems: an overview

2019

View full text Add to dashboard Cite

show abstract

“…While SuperBE as an algorithm is more complex and therefore slower than GMM or similar methods previously accelerated, it has an average PWC of 1.75%, much better than the 4% error rate expected from GMM (both scores measured on the CDW2014 dataset). In our work, we target a hard CPU with attached FPGA fabric, using a similar strategy from [10] to partition SuperBE into software and hardware components with the intention of accelerating computation on an embedded system. This improves upon the existing literature by accelerating a new algorithm that achieves better accuracy than most of the existing hardware implementations of background estimation algorithms, with real-time speeds on an embedded system.…”

Section: Hardware Implementationsmentioning

confidence: 99%

“…In a HW/SW Co-design system, some data communication has to occur between the Hardware (HW) component and the Software (SW) component, which requires a non-zero amount of time. In [10], we identified that the communication channels can become the bottleneck that prevents faster speeds from being achieved. If multiple non-sequential tasks are partitioned onto the hardware, then data needs to be passed between the HW and SW units multiple times.…”

Section: Hardware Accelerationmentioning

confidence: 99%

“…We described our hardware components in VHDL, and then synthesised onto the FPGA fabric. The Gaussian blur was implemented using a sliding window approach based on [10,36], shown in Figure 7. Essentially, the filter operation is parallelised so that the convolution operator can be applied to N × N pixels simultaneously, with a Gaussian kernel to create a blurring effect.…”

Section: Hardware Accelerationmentioning

confidence: 99%

“…Firstly, since the original SuperBE algorithm used RGB images with floating-point mathematics, we target software acceleration by investigating the effect of reducing the amount of information being processed by using greyscale and integer-only versions of SuperBE. Secondly, we developed an embedded system implementation of the algorithm with constrained computational resources for a real-world use case, targeting hardware acceleration by using Hardware/Software Co-design techniques [10] to partition the algorithm on a System-on-Chip (SoC) with an ARM processor and connected Field Programmable Gate Array (FPGA) fabric. In both cases, we present quantitative results justifying our acceleration strategies, while also detailing the effects on accuracy.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Accelerating SuperBE with Hardware/Software Co-Design

et al. 2018

Self Cite

View full text Add to dashboard Cite

Background Estimation is a common computer vision task, used for segmenting moving objects in video streams. This can be useful as a pre-processing step, isolating regions of interest for more complicated algorithms performing detection, recognition, and identification tasks, in order to reduce overall computation time. This is especially important in the context of embedded systems like smart cameras, which may need to process images with constrained computational resources. This work focuses on accelerating SuperBE, a superpixel-based background estimation algorithm that was designed for simplicity and reducing computational complexity while maintaining state-of-the-art levels of accuracy. We explore both software and hardware acceleration opportunities, converting the original algorithm into a greyscale, integer-only version, and using Hardware/Software Co-design to develop hardware acceleration components on FPGA fabric that assist a software processor. We achieved a 4.4× speed improvement with the software optimisations alone, and a 2× speed improvement with the hardware optimisations alone. When combined, these led to a 9× speed improvement on a Cyclone V System-on-Chip, delivering almost 38 fps on 320 × 240 resolution images.

show abstract

Violence Detection in Videos by Combining 3D Convolutional Neural Networks and Support Vector Machines

Accattoli

Sernani

Falcionelli

et al. 2020

Applied Artificial Intelligence

View full text Add to dashboard Cite

Video-surveillance has always been a vital tool to enforce safety in both public and private environments. Even though (smart) cameras are nowadays relatively widespread and cheap, such monitoring systems lack effectiveness in most scenarios. In addition, there is no guarantee about a human operator who monitors rare events in live video footages, forcing the use of such systems after unwanted events already took their undisturbed course, as a mere tool for investigations. Having an intelligent software to perform the task would allow to unlock the full potential of video-surveillance systems. To this end, in this paper we propose a solution based on a 3D Convolutional Neural Network that can effectively detect fights, aggressive motions and violence scenes in live video streams. Compared to state-of-the-art techniques, our method showed very promising performance on three challenging benchmark datasets: Hockey Fight, Crowd Violence and Movie Violence.

show abstract

Convolutional neural network acceleration with hardware/software co-design

Cited by 8 publications

References 38 publications

Computational workload in biometric identification systems: an overview

Computational workload in biometric identification systems: an overview

Accelerating SuperBE with Hardware/Software Co-Design

Violence Detection in Videos by Combining 3D Convolutional Neural Networks and Support Vector Machines

Contact Info

Product

Resources

About