2024 April ; 19(4): 471–478. | Hanyu Zheng, Quan Liu, Ivan I. Kravchenko, Xiaomeng Zhang, Yuankai Huo, Jason G. Valentine
The paper presents a meta-imager designed to accelerate machine vision applications by offloading computationally expensive convolution operations to high-speed, low-power optics. The meta-imager uses metasurfaces to enable angle and polarization multiplexing, allowing for multiple information channels to perform positive and negative convolution operations in a single shot. This approach is demonstrated for object classification tasks, achieving 98.6% accuracy in handwritten digits and 88.8% accuracy in fashion images. The meta-imager's compact size, high speed, and low power consumption make it suitable for a wide range of applications in artificial intelligence and machine vision. The design and fabrication of the meta-imager, as well as its performance in object classification, are detailed, showing that it can significantly reduce computational requirements and energy consumption.The paper presents a meta-imager designed to accelerate machine vision applications by offloading computationally expensive convolution operations to high-speed, low-power optics. The meta-imager uses metasurfaces to enable angle and polarization multiplexing, allowing for multiple information channels to perform positive and negative convolution operations in a single shot. This approach is demonstrated for object classification tasks, achieving 98.6% accuracy in handwritten digits and 88.8% accuracy in fashion images. The meta-imager's compact size, high speed, and low power consumption make it suitable for a wide range of applications in artificial intelligence and machine vision. The design and fabrication of the meta-imager, as well as its performance in object classification, are detailed, showing that it can significantly reduce computational requirements and energy consumption.