Neural bolometers: Designing next-generation infrared thermal imagers with in-pixel neuromorphic computing

Thermal imaging in the long-wave infrared (LWIR) spectrum is crucial for surveillance, navigation, and medical diagnostics. However, the latency and power consumption of current infrared detectors are stumbling blocks for efficient heat-based object detection, recognition, and ranging. In this paper, we present the concept of a neural bolometer that integrates the recently developed spintronic ultrafast nanoscale bolometers with an artificial neural network architecture. Our approach enables long-wave infrared detectors to function as both imaging sensors and stochastic computing units. Our design significantly enhances processing speed and reduces power consumption, achieving LWIR MNIST (Modified National Institute of Standards and Technology) classification with an inference time of 2.82 $\text{\ensuremath{\mu}}\mathrm{s}$ and power consumption of 1.9 $\text{\ensuremath{\mu}}\mathrm{W}$, while maintaining a high accuracy of 92.5%. The proposed approach can usher in a new generation of thermal imagers with in-pixel computing for widespread applications in defense, surveillance, healthcare, and autonomous navigation.