Introduction

This repository presents my approach to tackling the Airbus Ship Detection Challenge, employing two key models: a ship detection model and a ship segmentation model. I used two models for this, ship detection model and ship segmentation model. I decided to use two models instead of one driven by the fact that only 35% of the training images include ships, there's a considerable portion of the dataset where ships are absent, making segmentation challenging.

Segmentation model

Let's explore the ship segmentation model, which employs the U-Net architecture. It consists of six downsampling blocks(Conv+Conv+MaxPool) with number of filters [8, 16, 32, ..., 256] and five upsampling blocks(TransposeConv+ Concatenate+Conv+Conv). The output layer is Convolution with 1x1 filter and sigmoid activation. I chose a low number of filters because I lack the ability to train the model on powerful hardware. Additionally, ships are not geometrically complex objects, so a large number of filters is unnecessary for detecting their features.

Detection model

Now turning to the ship detection model, it follows a traditional Convolutional Network approach. It consists of three Conv+MaxPool blocks, along with one fully connected Dense layer and an output Dense layer with a sigmoid activation function.

Project Structure Overview

Directories

• ship_detection/: Contains files for creating dataset, training and evaluating ship detection model.
• ship_segmentation/: Contains files for RLE decoding, training, upgrading and evaluating ship detection model.

File Descriptions

• model_evaluation.py: Models evaluation
• eda-for-ship-detection.ipynb: File with exploratory data analysis
• dataset_for_detection.py: Script for ship detection dataset creation
• ship_detection_model.py: Script for creating and training detection model
• ship_detection_model.zip: Compressed ship detection model
• RLE_decoding: Script for decoding RLE string and creating dataset for ship segmentation model
• segmentation_model.py: Script for creating and training segmentation model
• ship_segmentation_model.keras: U-Net Segmentation model
• data_augmentation.py: Code for creating new dataset by using data augmentation
• segmentation_upgrade.py: Upgrading segmentation model by using new dataset and loss function
• ship_segmentation_model_v2.keras: New segmentation model
• what's_done.txt: File where I wrote all my progress while doing thing project
• resourses.txt: File containing articles and examples from which I sourced the information
• requirements.txt: File with required Python modules

Training the models

For both model fitting and dataset creating I used Google CLoud VM with Nvidia Tesla T4 GPU.

Loss functions:

• Segmentation model: Utilized the custom function dice_p_bce(), which combines binary cross-entropy and Dice coefficient
• Detection model: Employed the built-in binary_crossentropy() loss function

For training the segmentation model, I used a dataset with 42k pairs of images and masks, with each image containing at least one ship. Upon validation, the model achieved an accuracy of 0.75 as measured by the Dice coefficient.
For the ship detection model, I employed a training dataset consisting of 50,000 samples, complemented by a validation dataset of 10,000 samples. Continuing the evaluation, the model demonstrated a validation accuracy of 0.8534.