Author Affiliations
1Department of Optoelectronic Engineering, Jinan University, Guangzhou 510632, China2School of Optoelectronic Engineering, Guangdong Polytechnic Normal University, Guangzhou 510665, China3Guangdong Provincial Key Laboratory of Optical Fiber Sensing and Communications, Guangzhou 510632, Chinashow less
Fig. 1. Optical configuration of structured detected single-pixel imaging
Fig. 2. Framework of the fully convolutional neural network
Fig. 3. Optical-electronical hybrid neural network
Fig. 4. Example of the original training images and corresponding images with random rotation and lateral shift
Fig. 5. Confusion matrix of the classification results on handwritten digit test set (15 kernels)
Fig. 6. 2D convolutional kernel images of the first layer in the fully convolutional neural network
Fig. 7. MNIST test set classification accuracy of networks with different number of convolutional kernels
Fig. 8. Optical system. (a) Experimental setup; (b) Layout of the handwritten digits on disk
Fig. 9. A pair of binarized convolutional kernel images
Fig. 10. Snapshots of digit "5" in motion at different speeds captured by using a camera
Fig. 11. Single-pixel measurements of moving handwritten digits. (a) Single-pixel measurements of handwritten digits passing through the field of view successively in 1.5 s; (b) Partially enlarged view of the single-pixel measurements of the digit "5" in (a); (c) Result of the differential measurement from (b)
Fig. 12. The ten classes and example images in Fashion-MINST dataset
Fig. 13. Fashion-MINST test set classification accuracy of networks with different number of convolutional kernels
Linear velocity/m·s−1 | Number of kernels | Correct | Total | Correct/Total | 1.364 | 5 | 785 | 2181 | 35.99% | 10 | 523 | 681 | 76.80% | 15 | 584 | 607 | 96.21% | 20 | 339 | 339 | 100.00% | 25 | 323 | 346 | 93.35% | 30 | 180 | 195 | 92.31% | 2.450 | 5 | 737 | 2110 | 34.93% | 10 | 399 | 605 | 65.95% | 15 | 464 | 535 | 86.73% | 20 | 249 | 271 | 91.88% | 25 | 209 | 263 | 79.47% | 30 | 190 | 287 | 66.20% | 4.926 | 5 | 892 | 2679 | 33.30% | 10 | 543 | 973 | 55.81% | 15 | 420 | 625 | 67.20% | 20 | 190 | 332 | 57.23% | 25 | 145 | 326 | 44.48% | 30 | 114 | 301 | 37.87% |
|
Table 1. Experiment classification results of moving handwritten digits
Classifier name | Accuracy | Linear classifier [20] | 88.00% | SVM [23] | 98.60% | 6-layer neural network [24] | 99.65% | Deep convolutional network [25] | 99.65% | Proposed network | 97.99% |
|
Table 2. Results of different models on MNIST datasets