Fig. 6: Grayscale image and handwritten digits recognition using SMPD with different sampling masks.
From: Speckle-driven single-shot orbital angular momentum recognition with ultra-low sampling density
a Simulated scattering transmission and sparse detection: Images from the MNIST (handwritten digits) and Fashion-MNIST (clothing) datasets (28 × 28 pixels) were numerically propagating through a scattering medium (random transmission matrix), generating speckle patterns (256 × 256 pixels). Programmable sampling masks (7 × 7, 6 × 6, or 5 × 5 SPD arrays) extracted mean intensity values from 10 × 10 px regions, emulating sparse single-pixel detection. b Performance metrics: Recognition accuracy for each SPD configuration on the MNIST and Fashion-MNIST test datasets (6000 samples each). For MNIST, SMPD achieved recognition accuracies of 92.00%, 89.45%, and 83.28% with 7 × 7, 6 × 6, and 5 × 5 arrays, respectively. For Fashion-MNIST, the corresponding accuracies are 83.98%, 81.93%, and 81.17%, respectively. c Confusion matrices: Classification results for MNIST and Fashion-MNIST datasets using the 7 × 7 SPD array (49 detectors). Abbreviations: Cat., category; Acc., accuracy.
