The validation loss and validation accuracy for U-net are shown on the right side.
Testing = 0.0153
Test accuracy = 0.9272
In the test case, the accuracy of u-net reaches 93%, which is a good performance.
Obviously, as the training progresses, the loss gradually decreases and eventually converges.
In addition, processing one image only takes ~8 seconds on personal computers
Validation loss for U-net
Validation accuracy for U-net
The validation loss and validation accuracy simple CNN are shown on the right side.
Testing loss = 0.0543
Test accuracy = 0.9153
As the training progresses, the loss also gradually decreases and eventually converge. But there are some fluctuations in the process, not as smooth as U-net model.
Validation loss for simple CNN
Validation accuracy for simple CNN
The test sample of the u-net model is on the right side.
We could see that the prediction of u-net model is almost the same as the ground truth.
Input
Ground Truth
Output
In a concussion, comparing the results of U-net and simple CNN, we could find that the u-net model has better performance in both loss and accuracy.
In addition, processing time takes about 8 seconds on our personal computers. Comparing with the entropy-assisted kriging method that takes about 12 minutes, segmentation methods based on machine learning are much faster than typical methods while keeping high accuracy.
So, for Image Segmentation for predicting rock’s transport properties, machine learning based on training data generated by traditional ways can be used to process large volumes of images in a much shorter time.