Pixel-level pavement crack segmentation from UAV remote sensing images using ConvNeXt-UPerNet architecture. Implements the methodology from "Pixel-level pavement crack segmentation using UAV remote sensing images based on the ConvNeXt-UPerNet" (Taha et al., 2025).
| Component | Details |
|---|---|
| Encoder | ConvNeXt (Tiny/Small/Base) with ImageNet pretrained weights |
| Decoder | UPerNet with FPN + Pyramid Pooling Module |
| Input | 400x400x3 RGB images |
| Output | 400x400x1 binary segmentation masks |
| Loss | Combined Dice + Focal Loss (handles severe class imbalance) |
crack_segmentation_project/
├── model.py # ConvNeXt-UPerNet architecture
├── dataset.py # Dataset loading (supports .mat masks)
├── losses.py # Loss functions (Dice, Focal, Tversky, Combined, Boundary)
├── train_convnext.py # Training pipeline with mixed precision
├── evaluate_convnext.py # Evaluation with comprehensive metrics
├── predict_convnext.py # Inference with TTA support
├── models/
│ └── convnext.py # Standalone ConvNeXt implementation
├── utils/
│ ├── augmentation.py # Albumentations-based augmentation
│ └── patching.py # Image patching and preprocessing
├── data/
│ ├── images/ # Input UAV images (118 samples)
│ └── masks/ # Ground truth masks (.mat format)
├── evaluation_results/ # Evaluation metrics and plots
├── results/ # Prediction visualizations
├── plots/ # Training curves
└── logs/ # Training logs
git clone https://github.com/aliyanz85/cracks-detection-.git
cd cracks-detection-
pip install -r requirements.txtpython train_convnext.py \
--data_path data \
--backbone convnext_tiny \
--epochs 200 \
--batch_size 8 \
--learning_rate 1e-4python evaluate_convnext.py \
--model checkpoints/best.pth \
--data data \
--backbone convnext_tiny# Single image
python predict_convnext.py \
--model checkpoints/best.pth \
--input path/to/image.jpg \
--output predictions/
# Directory of images
python predict_convnext.py \
--model checkpoints/best.pth \
--input path/to/images/ \
--output predictions/ \
--tta # Enable test-time augmentation- Mixed Precision Training — Automatic FP16 for faster training on GPU
- Advanced Augmentation — Geometric + photometric transforms via Albumentations
- Test-Time Augmentation — Multi-flip ensemble for improved inference
- Multiple Loss Functions — Dice, Focal, Tversky, Combined, Boundary-aware
- Comprehensive Evaluation — IoU, Dice, F1, Precision, Recall, Confusion Matrix
- Flexible Backbone — ConvNeXt Tiny/Small/Base variants
| Parameter | Value |
|---|---|
| Optimizer | AdamW (β₁=0.9, β₂=0.999) |
| Learning Rate | 1e-4 with cosine annealing |
| Weight Decay | 1e-4 |
| Batch Size | 8 |
| Image Size | 400x400 |
| Epochs | 200 |
| Augmentation | Flip, Rotate, Brightness, Contrast, Noise, Blur |
The project uses the DronePavSeg UAV pavement crack dataset:
- 118 high-resolution UAV images
- Ground truth segmentation masks in MATLAB
.matformat - Binary labels: background (1) → 0, crack (2) → 1
Evaluation on the test set:
| Metric | Score |
|---|---|
| Pixel Accuracy | 96.7% |
| Precision | 26.7% |
| Recall | 58.9% |
| F1-Score | 36.7% |
Note: The low IoU/F1 reflects the extreme class imbalance in crack segmentation (cracks occupy <3% of pixels). Pixel accuracy is high due to the dominant background class. Further training with more epochs and hyperparameter tuning is recommended.
- Python 3.9+
- PyTorch 1.12+
- CUDA-capable GPU (recommended, 8GB+ VRAM)
- See requirements.txt for full dependencies
- Taha, M. et al. (2025). Pixel-level pavement crack segmentation using UAV remote sensing images based on the ConvNeXt-UPerNet.
- Liu, Z. et al. (2022). A ConvNet for the 2020s. CVPR.
- Xiao, T. et al. (2018). Unified Perceptual Parsing for Scene Understanding. ECCV.
MIT License