Advancing Low-Power Computer Vision at the Edge

Principal Investigator: Yung-Hsiang Lu

Learning-based computer vision requires vast amounts of computation passing information through many layers of neural networks. This project investigates how to improve energy efficiency of computer vision so that it can run on embedded systems, thus at the cameras, without transmitting video streams through networks. State-of-the-art computer vision aims to create general-purpose solutions that can recognize hundreds of classes of objects. In constract, this project creates much smaller neural networks that specialize on distingushing only a few types of objects. If more types need to be distinguished, this project creates hierarchies of neural networks and allow early termination when detected objects are not of interest. This project has already demonstrated significant reduction in energy consumption with negligible loss of accuracy.

Personnel

Other PIs: James Davis

Other Faculty: Neil Klingensmith and George Thiruvathukal (Loyola University Chicago)

Students: Abhinav Goel

Representative Publications

• "Low-Power Computer Vision: Status, Challenges, Opportunities", IEEE Journal on Emerging and Selected Topics in Circuits and Systems. Volume: 9 , Issue: 2 , June 2019

  "Low-power image recognition", Nature Machine Learning Vol 1, April 2019, 

   "Low-Power Image Recognition Challenge", AI Magazine Vol 39 No 2,  Summer 2018 

   

Keywords: computer vision, embedded system, energy efficiency, low power