Figure 6
Reconstruction of spatially overlapping phase images using a diffractive optical front-end (encoder) and a separately trained, shallow electronic neural network (decoder) with 2 hidden layers. The front-end diffractive optical networks are (a) D2NN-D1, (b) D2NN-D1d, (c) D2NN-D2, and (d) D2NN-D2d, shown in Figs. 2a, 3a, 4a and 5a, respectively. The detector layouts at the output plane of these diffractive optical networks are (a) D-1, (b) D-1d, (c) D-2, and (d) D-2d with \(2M\), \(4M\), \(M + 2\) and \(2M + 2\) single pixel detectors as shown in Fig. 1b-d, respectively; for handwritten digits \(M = 10\). These four designs create a compression ratio of 39.2 × , 19.6 × , 65.33 × and 35.63 × between the input and output fields-of-view of the corresponding diffractive optical network, respectively. The mean SSIM and PSNR values achieved by these phase image reconstruction networks are depicted in Table 2 along with the corresponding standard deviation values computed over the 10 K test input images (T2).
