Use some DNN and CNN networks to classify spectra with different signal-to-noise ratios
Use VGG16 net and res-18 net(change it to one-dimensional)to classify spectra with different signal-to-noise ratios
1.We found that in the case of limited information, the deeper network is not necessarily able to achieve satisfactory results, on the contrary, some simple networks can get unexpected surprises.
2.For spectral classification, high SN data often leads to data overfitting and poor generalization ability.
3.Compared with the simple network, the sensitivity of the structure to SNR is lower, and the amount of information extracted by the network is negatively correlated with the sensitivity.
If you have some ideas, please contact me [email protected].
SDSS DR14 data sets of m0-m4 with different SNR were used in the experiment
You can download my dataset from baidu drive.
| type | sn | number |
|---|---|---|
| M0 | 5-10 | 2850 |
| M0 | 10-15 | 1503 |
| M0 | up 15 | 2023 |
| M1 | 5-10 | 1919 |
| M1 | 10-15 | 1025 |
| M1 | up 15 | 1134 |
| M2 | 5-10 | 3300 |
| M2 | 10-15 | 1745 |
| M2 | up 15 | 1343 |
| M3 | 5-10 | 3105 |
| M3 | 10-15 | 1055 |
| M3 | up 15 | 1170 |
| M4 | 5-10 | 1658 |
| M4 | 10-15 | 779 |
| M4 | up 15 | 603 |
DNN
| sn | train acc | test acc |
|---|---|---|
| 5-10 | 89.5% | 85.7% |
| 10-15 | 85.6% | 87.1% |
| up15 | 87.6% | 90.3% |
CNN network
| sn | train acc | test acc |
|---|---|---|
| 5-10 | 97.5% | 90.8% |
| 10-15 | 96.6% | 94.7% |
| up15 | 93.4% | 75.4% |
vgg16
| sn | train acc | test acc |
|---|---|---|
| 5-10 | 99.0% | 89.0% |
| 10-15 | 97.8% | 90.9% |
| up15 | 96.9% | 66.9% |
res18
| sn | train acc | test acc |
|---|---|---|
| 5-10 | 98.9% | 75.2% |