Image-Segmentation---Research-Project

Project Structure

This project comprises several key components distributed in different folders for a clear organization and easy replication of experiments.

Tried Architectures (essaie_architecture)

We have tested various neural network architectures for our image segmentation task. To run a specific architecture, simply rename the desired folder to architecture.

The available architectures include:

• ResNet18
• ResNet101
• DenseNet121
• FCN (Fully Convolutional Network)
• EfficientNet (versions b0 and b1 with PPM module)

Loss Function Trials (essaieloss)

Several loss functions were experimented with to refine our model. To test them, rename the chosen loss function folder to loss.

The loss functions explored are:

• Binary Cross-Entropy (BCE)
• Dice Loss
• Focal Loss
• Boundary Loss
• Generalised Dice Loss
• Combination of BCE and Dice Loss

Baseline Model (best_model)

The best_model folder contains the model that achieved the best results during our tests. You can use it as a benchmark for performance.

Training Loss Visualization (plot_loss_exemple)

The plot_loss_exemple folder contains scripts for visualizing training loss curves. This can be useful for analyzing model behavior during training.

Usage

TTo run the segmentation model, use the following command:

python -m src.main --train

Before running the segmentation script, ensure that you have renamed the required architecture and loss function files. Simply name the architecture file you want to use as architecture.py, and the loss function file as loss.py. The currently employed architecture.py and loss.py represent the optimal combination, which is EfficientNet b1 + PPM, with the loss function being a combination of BCE + Dice. There's no need to change any other parts; just execute python -m src.main --train, and you will find the results in the submission folder within the baseline directory, which should contain many .npy files.

Additionally, if you wish to visualize the training loss, you can uncomment the last few lines in the final section of the script in the instance segmentation . This will generate and save an image of the loss curve.

If you wish to compare the predicted data with the actual data, you can use the command:

python -m src.main --evaluate

This will return an mIoU score, which is a standard metric for evaluating the quality of image segmentation models.

Acknowledgements

This project is built upon the models and frameworks provided at EY.