Warsaw University of Life Sciences

WC cemented carbides are widely used as tool materials for machining wood and wood-based products; however, a detailed microstructural characterisation-especially the distribution and size of WC grains-still depends on time-consuming, largely manual SEM image analysis. This work introduces a digital microstructure analysis pipeline in which the semi-automated segmentation and grain-measurement workflow is supported by accompanying scripts and datasets provided in the open repository (including training data generation, model checkpoints, and inference utilities), enabling reproducible grain boundary prediction and quantitative grain size extraction from SEM cross-section images.

08k_1100_003.png_458_220.png - is example of image-mask (training sample) pair (from approved-all.zip)

A single training sample consists of a cropped SEM image patch and an assigned binary mask of the same spatial size: the image shows WC grains with the background, and the mask contains pixels marking the grain area/fragment (black in the grain area) and the background (white outside the grain), thanks to which the model learns to predict the grain boundaries and shape based on the intensity in the image.

Files are:

• conda_env.txt - Conda environment specification used to run the scripts.
• S1. blobber.cpp - Computer-vision helper for refining labeled regions/masks (post-processing of annotation outputs).
• S1. generate_data.cpp - Generates training pairs (image-mask) from labeled data, including data augmentation.
• S1. model.pt - Preliminary model checkpoint produced after Stage 1 training.
• S1. pointer.cpp - Tool for initial polygon-based ROI/region refinement during manual annotation (edge-assisted refinement support).
• S1. test_unet_centered.py - Runs inference with the preliminary U-Net model to produce initial predictions for evaluation.
• S1. train_unet_centered.py - Trains a preliminary centered U-Net segmentation model using the generated image-mask dataset.
• S2. point_selector2.cpp - Improved manual annotation tool for selecting points used in the Stage 2 labeling workflow.
• S2. scaler.cpp - Image resizing utility used to scale images to/from the specified resolution (e.g., ~20k magnification).
• S2. smart_marker.cpp - Interactive (automatic + manual correction) grain boundary annotation tool; supports switching to manual mode.
• S2. viewer.cpp - Quality-control utility that classifies samples as approved or rejected based on expert review.
• S3. approved-all.zip - Curated/filtered training dataset containing only approved samples after Stage 2 QC.
• S3. filter_npz.py - Preprocessing script that computes edge-/gradient-based feature maps for each input image and saves them into per-image .npz files.
• S3. model v21, epoch_39.pth - Final selected model checkpoint chosen based on best model effectiveness coefficients.
• S3. train v21.py - Main training script for Stage 3 (version v21), including augmentation and edge-mask-based training.
• S3. Train.zip - Training dataset/package used for Stage 3 experiments.