BubbleSeg

BubbleSeg is a DeepLabV3+-based semantic segmentation framework for identifying bubble and solid phases in fluidized-bed images.

Features

• DeepLab-based segmentation with configurable backbone (default: ResNet18)
• VOC-style dataset loading and augmentation pipeline
• Multi-scale and sliding-window inference modes
• Automatic export of segmentation masks and overlay visualizations

Project Structure

BubbleSeg/
|-- base/
|-- data/
|   `-- BubbleDataSet/
|       |-- JPEGImages/
|       |-- SegmentationClass/
|       `-- ImageSets/
|           `-- Segmentation/
|-- dataloaders/
|-- models/
|-- pth/
|   |-- best_model.pth
|   `-- config.json
|-- test/
|-- config.json
|-- train.py
`-- inference.py

Dataset

The BubbleDataSet used in this project is publicly available on Zenodo:

Place the dataset under ./data/BubbleDataSet in VOC format (JPEGImages, SegmentationClass, and ImageSets/Segmentation).

If you use this dataset in research or production, please cite the DOI and follow the license and usage terms on the Zenodo page.

Requirements

• Python 3.8+
• PyTorch
• torchvision
• numpy
• scipy
• Pillow
• tqdm
• opencv-python

Example installation:

pip install torch torchvision numpy scipy pillow tqdm opencv-python

Training

Train with the project config:

python train.py --config config.json

Resume from a checkpoint:

python train.py --config config.json --resume path/to/checkpoint.pth

Select specific GPU devices (optional):

python train.py --config config.json --device 0

Inference

Run inference with the released checkpoint:

python inference.py --config ./pth/config.json --model ./pth/best_model.pth --test ./test --output outputs --extension jpg

Switch inference mode (optional):

python inference.py --config ./pth/config.json --model ./pth/best_model.pth --test ./test --mode sliding

Output

Inference results are saved to:

• outputs/masks: colorized segmentation masks
• outputs/overlays: side-by-side original image and mask overlay

Acknowledgement

This project is developed based on the open-source repository pytorch-segmentation by yassouali:
https://github.com/yassouali/pytorch-segmentation

Parts of the source code are simplified and modified for bubble/solid phase segmentation in fluidized-bed systems.

Contact

ybwj990122@163.com