In machine-vision applications, a front-end lens, also known as a meta-imager, developed at Vanderbilt University has the potential to replace conventional imaging optics by producing images faster and with reduced power consumption.
By reducing the thickness of the optical lens through nanostructuring, lens material can be used as a meta-imager filter, allowing for more efficient front-end processing and information encoding. Together with a digital backend, the imagers are intended to transfer computationally demanding tasks to high-speed, low-power optics. The generated images may find extensive use in the government and defense sectors, as well as in security and medical systems.
In a publication published in Nature Nanotechnology, professor Jason Valentine of mechanical engineering and deputy director of the Vanderbilt Institute of Nanoscale Science and Engineering present their proof-of-concept meta-imager.
Ivan I. Kravchenko, senior R&D staff member at the Center for Nanophase Materials Sciences, Oak Ridge National Laboratory, Quan Liu, a Ph.D. student in computer science, Hanyu Zheng, Ph.D., currently a postdoctoral associate at MIT, and assistant professor of computer science Yuankai Huo are among the other authors.
The authors point out that this architecture of a meta-imager can be extremely parallel and help close the gap between digital systems and the natural world. Valentine stated that there are numerous potential uses for this approach in machine vision, artificial intelligence, and information security because to its low power consumption, high speed, and small design.
An optic consisting of two metasurface lenses that encode data for a specific object classification task was first optimized by the team in order to create their meta-optic design. Networks trained on a handwritten number database and a clothes picture database frequently used to test different machine learning systems were utilized to create two fake versions. In handwritten digits and apparel photos, the meta-imager achieved 98.6% and 88.8% accuracy, respectively.
