shchukinvov / interf_seg

interf_seg

Interferograms snapshots segmentation.

Description

Basic information about interferometry and interferograms.

Example of interferogram is given below.

Main task

Build CV ML model to removing background from interferogram snapshot.

Minor task

Build classic CV algorithm using the obtained features of the model.

Validate metrics

• IoU score;
• Dice score;

Progress

1. Collected and annotated 212 grayscale images;

2. U-Net as the baseline model;

3. 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;

4. Best results at the moment:

   • IoUmean = 0.939
   • IoUmin = 0.776

   Received by following params:

   • Attention U-Net with features = (8, 16, 32, 64)
   • Adam optimizer
   • Learning rate: 4e-4
   • Batch size: 10
   • Num epochs: 450

Some images

TODO list

• Implement augmentations based on diffraction artifacts caused by dirt in optical system;