DisNet: deep learning approaches for dislocation segmentation in TEM

Models

• Implementation adapted from Xu et al. ArXiv (see repo Semicurv) and from Lan et al. ArXiv (see repo Elastic)

• Work with python 3.9 using pytorch. Conda environment details available in the environment.yml file.

• Training and inference scripts available in the Sbatch folder.

• Refer to main.py arguments for setting parameters

• Training results, weight and biases (.pt) available in the Results folder.

Synthetic

Generate the synthetic database

• See vtk_9.4.yml for requirements.

• xml_generator_calmip.py run the workflow. Need NUMODIS. Configuration files are in the generator folder.

• Use hw_numodis_interp_generator.py for image simulation (implementation from Eibl and Periano code; Microscopy and Microanalysis, 2007, 13 (S03),352-353)

DDPM

• Denoising Diffusion Probabilistic Model for background generation. Implementation from here. See ddpm_arm.yml for installation.

• Training and sampling from ddpm-train_gpu.py and ddpm-fond-sample.py scripts.

Human expert performance

Compare annotations of three images of different difficulties. compare-masks.py to compare between 2 users, save differences and compute IoU and CAL. compare-masks-all.py for pairwise comparison for all experts. Plot metrics histograms.

Inference on difficult images

• script_results_metrics_Disloc.py to compute evaluation metrics.

• plot_img.py and plot_hist.py to plot results comparison between models and plot metrics histograms.

Plotting results

Scripts for plotting comparison between predictions

Dislocation density measurement

Script to compute density from a segmented binary image.

Labeling

tutorial.pdf, tutorial for labeling images using GIMP. Script-Fu for GIMP used to import/export images and masks.

Datasets are available on request