Meta-optic Encoders For High-Speed Image Processing and Classification
Abstract
Convolutional neural networks that form the basis of artificial-intelligence and machine-learning systems have led to significant advances in image processing including edge detection and object recognition. However, the first stage layers in these networks contain large-scale computations. This results in the need for large computational resources, data rates, and power consumption, especially when processing large images and doing so in real-time. We propose to demonstrate high-speed edge detection and object classification by implementing multi-element meta-optic front-ends as passive encoders for digital convolutional neural networks. The meta-optic front-ends will serve to off-load the most computationally expensive tasks from the digital system, leading to an increase in speed and reduction in power consumption. These optical encoders operate at the speed of light and are completely passive. By using optics on the front-end we can also access additional information carriers, such as the angle of incidence, that can enable advanced processing using only a single optical layer. While the meta-optics will be passive, we will employ transfer learning architectures such that the digital back-end can be retrained for digital tasks. These concepts will be demonstrated in the long wave infrared (LWIR) though they can be scaled to arbitrary wavelengths. High-speed image processing, classification, and recognition, is critical for Naval supremacy as it allows for targets and threats to be identified in real-time. Off-loading large portions of the digital computation into passive ultra-thin optical elements will allow the Navy to reduce the size, weight, power consumption, and data rates of these systems while also increasing the speed of operation. The technology developed in this program can be engineered for a wide range of scenarios including target detection, threat detection, and navigation of autonomous vehicles pointing to application across many Naval platforms.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N000142112468
Entities
People
- Jason Valentine
Organizations
- Office of Naval Research
- United States Navy
- Vanderbilt University