Figure 2
All-optical classification of spatially-overlapping phase objects using the diffractive optical network D2NN-D1, based on the detector layout scheme (D-1) shown in Fig. 1b. (a) The thickness profiles of the diffractive layers constituting the diffractive optical network D2NN-D1 at the end of its training. This network achieves 82.70% blind inference accuracy on the test image set T2. (b–e) Top: Individual phase objects (examples) and the resulting input phase distribution created by their spatial overlap at the input field-of-view. Bottom: The normalized optical signals, \(I\), synthesized by D2NN-D1 at its output detectors. The output detectors with the largest 2 signals correctly reveal the classes of the overlapping input phase objects (indicated with the green rectangular frames).
