Fig. 1: A multichannel optical neural network (Monet) architecture for advanced machine vision tasks.
From: A multichannel optical computing architecture for advanced machine vision
a Network architecture of Monet and the projection-interference-prediction framework. Multiple observations are projected to a shared domain and encoded into optical fields, processed by interference units (IUs) for correspondence constructions and diffractive units (DUs) for feature embeddings. A regression module composed of iteratively connected IUs and DUs is adopted to predict the results for 3D perception or moving object detection. b Schematic and physical implementation of the IU (two-input). Multiple optical fields encoding multiple images are projected by task-specific function, propagate, and interfere to generate interference patterns. Different colours (red, green, blue) denote different visual inputs. In the physical implementation, two spatial light modulators are used to generate and project the optical fields, and a sensor is used to capture the interference pattern. c Schematic of the DU. A coherent light beam is modulated by SLM-1 to encode the input image, propagates to the SLM-2 for modulation, and propagates to the sensor for activation. DP diffractive propagation, BS beam splitter, Proj projection, Pred prediction
