In this repo I present my solution for the Airbus Ship Detection challenge on Kaggle. There was a data leakage in this competition which allowed for almost perfect scores. However, I personally used this comptetion as an introduction to computer vision and learned a lot about neural networks.
I developed two models. The fist is a quite simple image classification model which predicts if there is a boat within the image or not. The second is a sematic segmentation model which divides the image into background and foreground (=ships). Both neural networks are based on the Tensorflow wrapper Keras. In a third step the pictures which contain no boat are determined using the classification model and written to a dataframe. For the images that do contain a boat the instance segmentation model is run. Then the predicted masks gets encoded and are added to the dataframe. This dataframe contains the predictions for the entire test dataset.
For the image classification model I have trained a simple Convolution Neural Network (CNN). It takes an input of 512 x 512 x 3 image and generates an output of (None, 1).
This is the visualization of the network:
It consists of three convolution blocks. Each block contains two Convolution layers and one Max Pooling layer. The top of the network consists of one Global Max Pooling layer, a Dense and Dropout layer. During the running I managed to get an accurarcy of up to 89% with this fairly simple setup.
The instance segmentation model was inspired by TernausNet. An implementation of the U-net architecture can be found here. The architecture of the U-net looks like this:
In the paper the authors argue that a U-net architecture can be improved with a pre-trained encoder. The authors propose a VGG-11 encoder and that networks with a pre-trained encoder converge faster than networks without. However, since Keras offers a VGG16 pre-trained on the ImageNet dataset, I decided to use this encoder. The final architecture looks like this:
