Gas Pipe Detection

AI models for gas pipe detection and segmentation, developed as part of the EXPAND CoRe 1 project.

This repository trains and evaluates a deep learning model that segments gas pipes in images, then derives pipe-level detection metrics (accuracy, precision, recall, F1) on top of the raw pixel-level segmentation.

Overview

The core pipeline:

1. Loads a custom gas-pipe image dataset (GasPipeDataset) and splits it into train/validation/test sets.
2. Applies data augmentation (resize, flips, rotation, brightness/contrast jitter, Gaussian blur, normalization) via Albumentations.
3. Trains a U-Net segmentation model (src/models/unet.py) to predict pipe masks.
4. Evaluates the trained model across multiple decision thresholds and reports:
   • Segmentation metrics: Dice coefficient, IoU score, pixel accuracy
   • Detection metrics: detection accuracy, precision, recall, F1-score
   • Pixel classification metrics: pixel precision, recall, F1-score
5. Automatically selects the best threshold (by IoU) and generates a confusion matrix plus qualitative prediction visualizations.

Repository structure

core-pipe-detection/
├── configs/       # Experiment / model configuration files
├── data/          # Dataset location (see Data section)
├── src/
│   ├── datasets/
│   │   └── GasPipeDataset.py   # Custom PyTorch Dataset + train/val split helper
│   ├── models/
│   │   └── unet.py             # U-Net model, training loop, evaluation & plotting utilities
│   ├── logging.py              # Custom logging handler
│   └── parser.py               # Config/argument parser
├── requirements.txt
└── train.py       # Main entry point: train + evaluate the model

Requirements

Install everything with:

pip install -r requirements.txt

Note: A CUDA-capable GPU is strongly recommended for training.

Data

GasPipeDataset expects an image/mask dataset laid out under a single root directory (path configured via dataset.path in your config). The dataset loader automatically creates an 80/20 train/validation split (train_ratio=0.8, fixed random_state=42 for reproducibility). Place your data under the data/ directory, or point the config at your own dataset location.

Configuration

Runs are driven by a config file (see the configs/ folder) parsed via src/parser.py. At minimum, a config should define:

dataset:
  path: /path/to/dataset

model:
  batch_size: 8
  n_epochs: 50
  lr: 0.0001
  checkpoints: /path/to/checkpoints/model.pth

Usage

Train (and evaluate) a model with:

python train.py --config configs/<your_config>.yaml

What happens under the hood:

• If a checkpoint already exists at config.model.checkpoints, it is loaded; otherwise a new U-Net is created.
• The model is trained for config.model.n_epochs epochs, and training/validation loss curves plus Dice scores are plotted and saved next to the checkpoint (*_training_history.png).
• The best checkpoint is reloaded and evaluated across five thresholds (0.5–0.9); per-threshold results are saved to *_threshold_analysis.json.
• The threshold with the highest IoU is used to generate a confusion matrix (*_confusion_matrix.png) and a qualitative sample of predictions.

Outputs

For a given checkpoint path (e.g. checkpoints/model.pth), training produces:

File	Description
model.pth	Best model weights
model_training_history.png	Train/val loss and validation Dice score curves
model_threshold_analysis.json	Segmentation/detection/pixel metrics at each tested threshold
model_confusion_matrix.png	Confusion matrix at the best-performing threshold