🤖 AI Summary
Diffraction neural networks face a critical challenge in active illumination scenarios (e.g., microscopy, autonomous driving): when the object’s spatial coherence length becomes comparable to the system’s minimum resolvable feature size, conventional coherent or incoherent modeling extremes fail. This work systematically reveals, for the first time, the decisive impact of spatiotemporal coherence on network generalization. We propose the first coherence-aware, end-to-end differentiable modeling and joint optimization framework. Grounded in rigorous wave-optical theory, it enables complex-amplitude wave propagation simulation and gradient backpropagation for arbitrarily specified degrees of coherence, supporting both linear and nonlinear diffraction layers. Experiments demonstrate substantial improvements in classification accuracy across diverse coherence conditions, validating coherence as an essential and effective independent design degree of freedom.
📝 Abstract
We demonstrate the significant influence of the illumination coherence on diffractive networks, and propose a framework for network optimization with any prescribed degree of spatial and temporal coherence. We analyze performance for varied coherence properties.