Physics‐Inspired Diffractive Neural Networks for Zooming Image Edge Detection

Diffractive neural networks (DNNs) are emerging as a novel computing architecture for image processing due to their low computational power and high‐speed. However, existing DNN‐based image processors generally employ a pure data‐driven training strategy, and require a sufficiently large training set, suffering from poor generalization and interpretability. Moreover, these methods typically utilize multiple cascaded diffractive layers to improve the performance of networks, which may inevitably encounter challenges of cascaded alignment and diffraction efficiency. Here, a physics‐inspired DNN (PI‐DNN) is proposed and experimentally demonstrated for image edge detection. By embedding the DNN layers into the physical model of traditional lens imaging, PI‐DNN architecture requires only two diffractive layers to achieve zooming all‐optical edge detection tasks, exhibiting strong robustness and generalization. Thus, it not only enables few‐shot data training but also greatly reduces the number of required DNN layers, avoiding alignment difficulties and efficiency losses compared with regular DNNs. As a proof‐of‐concept, the PI‐DNN‐based edge detection tasks with scaling factors of 0.5, 1.0, and 1.5 on both intensity and phase images are demonstrated. The proposed framework opens up a new physics‐inspired paradigm for designing task‐specific all‐optical image processors, significantly accelerating their practical applications in biological microscopy and machine vision.