Segmentation of PMSE signals

A repository for segmentation of PMSE signal using UNet and UNet++ architecture.

The code originates from the work done in a Master's thesis and in a published paper.

Clone repository:

   git clone https://github.com/Domben93/PMSEsegmentation.git

Install dependencies:

The required libraries are found in /requirements.txt and can be installed using:

   pip install -r requirements.txt

Inference

Put the data samples into the test folder as shown in the structure below. The sample data is put in the data folder. NOTE that the data must be either .png or .mat extension files.

.
├── ...
├── Dataset
│   ├── ...    
│   ├── Test
│   │   └── data
│   └── ...                
└── ...

###Pretrained weights:

Pretrained weights can be downloaded from a shared google drive folder. Only the best performing weights are made available. See Train on how to train your own weights from scratch.

###Run inference:

• Quick Run

   python inference.py

• Specify image, label, model and model-wights paths. It

   python inference.py --images <image_path> --labels <label_path> --model <model_path> --weights <wight_path>

For additional inference specifications please see code.

Train

To train the model put the training samples into the Train folder and validation samples into the Validation folder. The sample data is put in the data folder and the labels in the label folder as shown in the structure below. NOTE that the data samples and labels must have the same name and must be either .png or .mat files.

.
├── ...
├── Dataset 
│   ├── Train
│   │   ├── data
│   │   └── label
│   ├── Validation
│   │   ├── data
│   │   └── label
│   └── ...                
└── ...

Customize config/unet_config.yaml with wanted parameters

Note that the augmentation must be selected/changed in the code.

For training run:

   python train.py

Training model with ObjectAug generated data

The code does not support simultaneously generation of new data with the ObjectAug method and at the same time of train a model. Instead, the data can be generated beforehand using /generate_data.py. If the data-path and save-path is not specified the path is set to the /Dataset/Train and /Dataset/Train/generated_data, respectively. These paths can easily be changed to the preferred directories.

The ObjectAug model uses a pretrained model for the inpainting process which employs a UNet model with partial convolution. The code for the inpainting model is copied from tanimutomo and slightly modified.

To train a new inpainting model from scratch please see /inpainting/README.md. NOTE that in the code from the inpainting model has been altered from the original git repository and might not function according to what the README.md says.

The weights used in the generation of the data used to train the models in the also be downloaded from a shared google drive folder.

To run the generate_data.py please see available parser arguments in the code, and run it using:

    python generate_data.py

Test

   python test.py

NOTE:

The code is not as clean or well documented as I would like and apologize for the mess that it may seem. I will try to update the code and documentation in the future but as of now if you have any questions feel free write.

Reference Codes

• https://github.com/tanimutomo/partialconv
• https://github.com/MrGiovanni/UNetPlusPlus

Original Thesis/Paper

• Thesis

  • Thesis Title:
    • Segmentation of Polar Mesospheric Summer Echoes using Fully Convolutional Network
  • Author:
    • Domben, Erik Seip
  • Affiliation:
    • UiT The Arctic University of Norway

• Paper

  • Paper Title:
    • Using Deep Learning Methods for Segmenting Polar Mesospheric Summer Echoes
  • Authors:
    • Domben, Erik Seip; Sharma, Puneet; Mann, Ingrid
  • Affiliation:
    • UiT The Arctic University of Norway