Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition

Sungho Park, Ahmed Al Maashri, Yang Xiao, Kevin M. Irick, Vijaykrishnan Narayanan

Research output: Chapter in Book/Report/Conference proceedingConference contribution

2 Citations (Scopus)

Abstract

Object recognition is one of the most important tasks in computer vision due to its wide variety of applications from small hand-held devices to surveillance systems in large public facilities. Even though biologically inspired approaches have been recently revealed to take another significant step forward to reduce its large power consumption, it still consumes relatively large amounts of energy because of the immense amount of data and computations. Typically in such biologically inspired - often called neuromorphic - object recognition implementations, visual saliency feeds feature extraction to limit the amount of computations effectively by picking a pre-determined size of patches around salient locations of an image. In this work, we explore the design space of HMAX for neuromorphic feature-extraction and classification along with the trade-off between energy consumption and classification accuracy. In addition, a novel method to further reduce energy consumption is proposed by leveraging effort-level of HMAX according to the findings of visual saliency in an efficient manner. Experiments revealed that our dynamic configuration achieved 70.57% of energy reduction with only 1.05% of accuracy loss for accuracy-critical applications. For energy-critical applications, a proposed configurations trades off 5.07% accuracy to gain 91.72% reduction in energy consumption.

Original languageEnglish
Title of host publicationProceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013
PublisherIEEE Computer Society
Pages139-144
Number of pages6
ISBN (Print)9781479913312
DOIs
Publication statusPublished - 2013
Event2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013 - Natal, Brazil
Duration: Aug 5 2013Aug 7 2013

Other

Other2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013
CountryBrazil
CityNatal
Period8/5/138/7/13

Fingerprint

Object recognition
Energy utilization
Feature extraction
Computer vision
Electric power utilization
Experiments

Keywords

  • dynamic configuration
  • energy efficiency
  • FPGA
  • HMAX
  • object recognition
  • visual saliency

ASJC Scopus subject areas

  • Hardware and Architecture
  • Control and Systems Engineering
  • Electrical and Electronic Engineering

Cite this

Park, S., Al Maashri, A., Xiao, Y., Irick, K. M., & Narayanan, V. (2013). Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition. In Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013 (pp. 139-144). [6654636] IEEE Computer Society. https://doi.org/10.1109/ISVLSI.2013.6654636

Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition. / Park, Sungho; Al Maashri, Ahmed; Xiao, Yang; Irick, Kevin M.; Narayanan, Vijaykrishnan.

Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013. IEEE Computer Society, 2013. p. 139-144 6654636.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Park, S, Al Maashri, A, Xiao, Y, Irick, KM & Narayanan, V 2013, Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition. in Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013., 6654636, IEEE Computer Society, pp. 139-144, 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013, Natal, Brazil, 8/5/13. https://doi.org/10.1109/ISVLSI.2013.6654636
Park S, Al Maashri A, Xiao Y, Irick KM, Narayanan V. Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition. In Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013. IEEE Computer Society. 2013. p. 139-144. 6654636 https://doi.org/10.1109/ISVLSI.2013.6654636
Park, Sungho ; Al Maashri, Ahmed ; Xiao, Yang ; Irick, Kevin M. ; Narayanan, Vijaykrishnan. / Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition. Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013. IEEE Computer Society, 2013. pp. 139-144
@inproceedings{dd88ee5ccc374884b17d4b940d7be92a,
title = "Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition",
abstract = "Object recognition is one of the most important tasks in computer vision due to its wide variety of applications from small hand-held devices to surveillance systems in large public facilities. Even though biologically inspired approaches have been recently revealed to take another significant step forward to reduce its large power consumption, it still consumes relatively large amounts of energy because of the immense amount of data and computations. Typically in such biologically inspired - often called neuromorphic - object recognition implementations, visual saliency feeds feature extraction to limit the amount of computations effectively by picking a pre-determined size of patches around salient locations of an image. In this work, we explore the design space of HMAX for neuromorphic feature-extraction and classification along with the trade-off between energy consumption and classification accuracy. In addition, a novel method to further reduce energy consumption is proposed by leveraging effort-level of HMAX according to the findings of visual saliency in an efficient manner. Experiments revealed that our dynamic configuration achieved 70.57{\%} of energy reduction with only 1.05{\%} of accuracy loss for accuracy-critical applications. For energy-critical applications, a proposed configurations trades off 5.07{\%} accuracy to gain 91.72{\%} reduction in energy consumption.",
keywords = "dynamic configuration, energy efficiency, FPGA, HMAX, object recognition, visual saliency",
author = "Sungho Park and {Al Maashri}, Ahmed and Yang Xiao and Irick, {Kevin M.} and Vijaykrishnan Narayanan",
year = "2013",
doi = "10.1109/ISVLSI.2013.6654636",
language = "English",
isbn = "9781479913312",
pages = "139--144",
booktitle = "Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013",
publisher = "IEEE Computer Society",

}

TY - GEN

T1 - Saliency-driven dynamic configuration of HMAX for energy-efficient multi-object recognition

AU - Park, Sungho

AU - Al Maashri, Ahmed

AU - Xiao, Yang

AU - Irick, Kevin M.

AU - Narayanan, Vijaykrishnan

PY - 2013

Y1 - 2013

N2 - Object recognition is one of the most important tasks in computer vision due to its wide variety of applications from small hand-held devices to surveillance systems in large public facilities. Even though biologically inspired approaches have been recently revealed to take another significant step forward to reduce its large power consumption, it still consumes relatively large amounts of energy because of the immense amount of data and computations. Typically in such biologically inspired - often called neuromorphic - object recognition implementations, visual saliency feeds feature extraction to limit the amount of computations effectively by picking a pre-determined size of patches around salient locations of an image. In this work, we explore the design space of HMAX for neuromorphic feature-extraction and classification along with the trade-off between energy consumption and classification accuracy. In addition, a novel method to further reduce energy consumption is proposed by leveraging effort-level of HMAX according to the findings of visual saliency in an efficient manner. Experiments revealed that our dynamic configuration achieved 70.57% of energy reduction with only 1.05% of accuracy loss for accuracy-critical applications. For energy-critical applications, a proposed configurations trades off 5.07% accuracy to gain 91.72% reduction in energy consumption.

AB - Object recognition is one of the most important tasks in computer vision due to its wide variety of applications from small hand-held devices to surveillance systems in large public facilities. Even though biologically inspired approaches have been recently revealed to take another significant step forward to reduce its large power consumption, it still consumes relatively large amounts of energy because of the immense amount of data and computations. Typically in such biologically inspired - often called neuromorphic - object recognition implementations, visual saliency feeds feature extraction to limit the amount of computations effectively by picking a pre-determined size of patches around salient locations of an image. In this work, we explore the design space of HMAX for neuromorphic feature-extraction and classification along with the trade-off between energy consumption and classification accuracy. In addition, a novel method to further reduce energy consumption is proposed by leveraging effort-level of HMAX according to the findings of visual saliency in an efficient manner. Experiments revealed that our dynamic configuration achieved 70.57% of energy reduction with only 1.05% of accuracy loss for accuracy-critical applications. For energy-critical applications, a proposed configurations trades off 5.07% accuracy to gain 91.72% reduction in energy consumption.

KW - dynamic configuration

KW - energy efficiency

KW - FPGA

KW - HMAX

KW - object recognition

KW - visual saliency

UR - http://www.scopus.com/inward/record.url?scp=84893614445&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84893614445&partnerID=8YFLogxK

U2 - 10.1109/ISVLSI.2013.6654636

DO - 10.1109/ISVLSI.2013.6654636

M3 - Conference contribution

SN - 9781479913312

SP - 139

EP - 144

BT - Proceedings - 2013 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2013

PB - IEEE Computer Society

ER -