asd_challenge

Airbus Ship Detection Challenge

---

About Project

That my implementation for Airbus Ship Detection Challenge. Dataset downloaded from Kaggle Challenge, you also can download it if you have Kaggle account. My model based on TensorFlow/Keras frameworks. Implementation containing model architecture such as a simple U-Net for image semantic segmentation. In project also used the libraries: NumPy, Pandas and Matplotlib.

Project have that files:

• dataset_analysis.ipynb -- containing exploring of data from Challenge.
• cfg/config.yaml -- for storing config variables for model.
• model_debugger.py -- it's core code file with calling all function and classes for preprocessing data, building, training and using model.
• model_builder.py -- file with function for building model with Keras functional API and classes of custom
• metric and loss function.
• training_process.py -- file with function for training model.
• inference_process.py -- file with function for testing model.
• data_handler.py -- file with class for building batch with images and masks.
• utils.py -- file with all support functions.
• requirements.txt -- file which store name all libraries that require for this project.

---

Model Architecture

Model architecture has:

• Input layer with shape=(160, 160)
• data_augmentation layer (for augmented data only in training process)

And standard U-Net architecture from code example in Keras documentation (Image Segmentation with a U-Net-like architecture):

My model build code for this project

---

Downloading and testing model

First get the repository

git clone https://github.com/lytmercy/asd_challenge.git

Next you need step into repo directory

cd asd_challenge

Next create new python environment (my python is 3.10).

python -m venv venv

Activate new environment

./venv/Scripts/activate

And install all requirements

pip install -r requirements.txt

And finally run the debugger.py in console

python src/model_debugger.py

If you want check training process you can move out files of weight from path models/trained/weights/*

mv models/trained/weights/* your/path/for/my/weights/*

And next run again model_debugger.py

---

Demonstrating the result

In order not to wait a long time for the training result, I used only 30% of the images from those with a ground-truth mask.

Dice Loss

For Dice Coefficient, I use this formula and subtract it from 1, to get dice loss. From this science paper (on page 6):

Hyperparameters

For the training I use that hyperparameters:

• learning_rate = 0.001
• number of epochs = 4
• batch size = 22 (because my GPU can't process more)

Results

In result of training, I got this:

• dice_loss = 0.1890
• dice_score = 0.8150
• val_dice_loss = 0.3031
• val_dice_score = 0.7109

And got this curves of loss and dice score:

Dice Loss Function

Dice Score Metric

In result of evaluating, I got this:

• dice_loss = 0.3924
• dice_score = 0.6135

Prediction from model

And take prediction from model, I got this:

Good results with one or more ships:

And bad results model predicts when getting an image with some terrain and some other stuff that recognize as ships on empty sea images.