journal menu
![[Figure 5]](mo5274fig5mag.jpg)
|
|
Figure 5
Analysis of the accuracy and error loss metrics for training and evaluation of ten Sensor3D and ten U-Net models. For most of the models, accuracy improves and the error decreases when increasing the number of training images. The main effect of the batch size is observed for models trained with a smaller training data size of 10 and 20, though the difference is not statistically significant. Whereas the Sensor3D model with a batch size of 32 and training size of 10, 20 and 30 required 71, 67 and 52 epochs, respectively, unexpectedly for a training size of 50, only 30 epochs were needed due to a lack of any increase in learning rate, whereas with 70 training images a total of 97 epochs were needed. For the Sensor3D model with a batch size of 64 and training sizes of 10, 20, 30, 50 and 70, the DL process required 65, 30, 45, 56 and 90 epochs, respectively, thus increasing with increasing training data sizes. The U-Net model with a batch size of 32 and a training sizes of 10, 20, 30, 50 and 70 images required 30, 43, 45, 56 and 51 epochs, respectively. The U-Net model with a batch size of 64 and training sizes of 10, 20, 30, 50 and 70 images were completed within 65, 25, 119, 55 and 65 epochs, respectively, again demonstrating that the number of epochs required is not linearly related to the training data size. The two intermediate additional data sizes of 10 and 30 ground-truth images are marked by gray shade regions. |
![[Figure 5]](mo5274fig5.jpg)
 | JOURNAL OF SYNCHROTRON RADIATION |
ISSN: 1600-5775
Open 
access
access
