Interferograms snapshots segmentation.
Basic information about interferometry and interferograms.
Example of interferogram is given below.
Build CV ML model to removing background from interferogram snapshot.
Build classic CV algorithm using the obtained features of the model.
- IoU score;
- Dice score;
-
Collected and annotated 212 grayscale images;
-
U-Net
as the baseline model; -
Tested following cases:
U-Net
,SegNet
,MSU-Net
,AttU-Net
models;BCELoss
,IoULoss
,ScaledIoULoss
,IoUWithBCELoss
losses;- Pretraining model on large amount of generated data;
- Postprocessing based on morphological transforms;
-
Best results at the moment:
IoUmean = 0.939
IoUmin = 0.776
Attention U-Net
withfeatures = (8, 16, 32, 64)
Adam
optimizer- Learning rate:
4e-4
- Batch size:
10
- Num epochs:
450
Some images
- Implement augmentations based on diffraction artifacts caused by dirt in optical system;