Review of optical pre-sensor computing

Fueled by the rapid progress of artificial intelligence, machine vision has evolved rapidly in recent years. However, current architectures still rely on cascaded imaging, sensing, transmission and processing stages, leading to substantial data movement, energy consumption and latency. Recent efforts to alleviate these bottlenecks have shifted computation progressively closer to the data source, from backend processors to near-sensor and in-sensor computing units. Optical pre-sensor computing extends this trend to an earlier stage by embedding computation directly into optical propagation before photodetection. It enables task-relevant visual information to be extracted in the optical domain, reducing redundant data acquisition and easing subsequent electronic processing. Benefiting from the intrinsic speed, bandwidth and parallelism of light, this emerging paradigm offers a promising route toward efficient front-end visual information processing. This paper provides a comprehensive review of recent progress in optical pre-sensor computing. Its system architectures and implementation frameworks are first outlined. Advances in free-space optical operators, optical neural networks, and nonlinear optical activation mechanisms are then summarized. Finally, key challenges and opportunities for visual perception in physical AI are discussed, and future directions for optical pre-sensor computing in next-generation machine vision systems are highlighted, with particular emphasis on low-power consumption, high-level integration, and reconfigurable architectures.