A meta-imager is introduced to accelerate machine vision by offloading computationally intensive convolution operations into high-speed, low-power optics. This device uses metasurfaces to enable both angle and polarization multiplexing, creating multiple information channels that perform positive and negative valued convolution operations in a single shot. The meta-imager achieves 98.6% accuracy in handwritten digit classification and 88.8% accuracy in fashion image classification. It is compact, fast, and energy-efficient, making it suitable for a wide range of AI and machine vision applications. The meta-imager is designed to replace traditional imaging optics in machine vision applications by encoding information in a more efficient basis for backend processing. It enables incoherent illumination and a wide field of view, essential for imaging natural scenes with ambient lighting. The device uses angular and polarization multiplexing to achieve multiple independent, arbitrary convolution channels. The first metasurface splits the incident signal into angular channels, while the second serves as a focusing optic to create an N×N focal-spot array for each channel. The meta-imager is fabricated using silicon nanopillars as the base meta-atoms, which serve as half-wave plates. The second metasurface is designed for polarization-insensitive phase control. The device is characterized using a linearly polarized laser to obtain the PSF, showing a good match between ideal and measured results. The meta-imager is tested with the MNIST and Fashion MNIST datasets, achieving high classification accuracy. The meta-imager's design allows for efficient optical convolution, with 94% of the operations offloaded from the digital platform into the front-end optics. This results in a significant speedup for classification tasks. The device is scalable, with classification accuracy as a function of the areal density of the basic computing unit. The system's functionality remains unchanged with up to six times higher areal computing unit density. The meta-imager has potential applications beyond machine vision, including information security and quantum communications. It offers a highly parallel architecture that bridges the gap between the natural world and digital systems. The device is a proof of concept for a convolutional front end that can replace traditional imaging optics in machine vision applications.A meta-imager is introduced to accelerate machine vision by offloading computationally intensive convolution operations into high-speed, low-power optics. This device uses metasurfaces to enable both angle and polarization multiplexing, creating multiple information channels that perform positive and negative valued convolution operations in a single shot. The meta-imager achieves 98.6% accuracy in handwritten digit classification and 88.8% accuracy in fashion image classification. It is compact, fast, and energy-efficient, making it suitable for a wide range of AI and machine vision applications. The meta-imager is designed to replace traditional imaging optics in machine vision applications by encoding information in a more efficient basis for backend processing. It enables incoherent illumination and a wide field of view, essential for imaging natural scenes with ambient lighting. The device uses angular and polarization multiplexing to achieve multiple independent, arbitrary convolution channels. The first metasurface splits the incident signal into angular channels, while the second serves as a focusing optic to create an N×N focal-spot array for each channel. The meta-imager is fabricated using silicon nanopillars as the base meta-atoms, which serve as half-wave plates. The second metasurface is designed for polarization-insensitive phase control. The device is characterized using a linearly polarized laser to obtain the PSF, showing a good match between ideal and measured results. The meta-imager is tested with the MNIST and Fashion MNIST datasets, achieving high classification accuracy. The meta-imager's design allows for efficient optical convolution, with 94% of the operations offloaded from the digital platform into the front-end optics. This results in a significant speedup for classification tasks. The device is scalable, with classification accuracy as a function of the areal density of the basic computing unit. The system's functionality remains unchanged with up to six times higher areal computing unit density. The meta-imager has potential applications beyond machine vision, including information security and quantum communications. It offers a highly parallel architecture that bridges the gap between the natural world and digital systems. The device is a proof of concept for a convolutional front end that can replace traditional imaging optics in machine vision applications.