Fig. 2
From: Optical neural networks with intensity‑based projection layers as effective nonlinear activations
Implementations of CVNN with complex projection layers on feature extraction tasks across three different levels. (a) Schematic of the network for image classification. (b)-(c) Testing results for classification on (b) MNIST dataset and (c) Fashion MNIST dataset, with the averaged accuracy of 97.49% and 89.11%, respectively. (d) Schematic of the network for image reconstruction. (e)-(f) Testing results for image reconstruction on (e) MNIST dataset and (f) Fashion MNIST dataset. (g) Schematic of the network for the specific image feature extraction. The input image is a white circle in a square box. The radius r and position (x, y) of the circle can be varied. By incorporating a digital linear layer (denoted as “L”) into the CVNN, the trained network is capable of extracting the three features embedded in the input image. (h) Testing results for image feature extraction.
