Academic researchers say YOLO, speed up image recognition tech
Academic researchers say they have managed to juggle two competing requirements for image recognition: improving both the accuracy and the speed of image recognition. The software is already being used in Smart City systems for traffic management but has relevance for speeding up biometric identification (including face mask detection) in real time.
The researchers at the Institute of Information Science Academia Sinica in Taiwan worked on improving existing image recognition software, dubbed YOLO (for You Only Look Once) with the current code maintainer Alexey Bochkovskiy. They worked on speeding up the underlying neural network by tuning “the response of network identification objects [and the] retransmission mechanism to optimize the transmission path, [and] to reduce the amount of calculation algorithms,” according to a news release.
The end result is that the researchers Wang Jianyao and Liao Hongyuan claim that the new version of YOLO reaches a higher accuracy rate that is 10 percent higher than the previous generation and is more accurate than other image recognition technologies currently in use, it is claimed.
The software is being used by Elan Electronics as part of a “smart city traffic flow solution” which is currently deployed in Taoyuan and Hsinchu in Taiwan. Image recognition is applied to vehicles passing through intersections in order to improve the ability to detect vehicles, parking trains and vehicle speeds in real time at each intersection.
Image recognition at the speed of traffic is no easy feat, which is why the use of the software for facial recognition in crowds for mask detection is one example of its usefulness for biometric identification purposes.
Another benefit of the new software: in their joint research paper, the team noted that the reduced complexity of calculations not only improves speed but means that training the algorithms requires less compute power and time and can be done on a variety of currently available GPUs.
The resulting software is available on open source code repository GitHub.