Sample Images: https://drive.google.com/open?id=1rjQF9jnoygp29ucw-GE5W1EPSuTt8L-g
This is a Kaggle challenge, here is the link. https://www.kaggle.com/c/airbusship-detection. The data size is about 30GB. The training data are masked with a separate csv file. We need to submit the same mask csv file for the test data.
We have developed an ensemble model, combining a pretrained resnet 50 and VGG16, to detect whether there is a ship in the image or not. The accuracy is about 95%. Considering the fact that users could upload lots of images in the same time and no ships on about 70% of the images, a simply model can speed up the whole process.
For ship-existence model, we refer to keras.application documents, https://keras.io/applications/. We use VGG16 and ResNet50 pretrained with ImageNet, add dense layer and flatten layer after the pretrained model. And combin in Model folder.Loss function is binary_crossentropy for we only need to detect if there is a ship or not. We tried simple CNN at first, but it doesn’t work well. Then we used pretrained models trained with imagenet. These models have been well trained for classification purpose, which saved us a lot of time.
We used a U-net model to predicte the location of the ships and draw a border-box around them.
For U-net model, referring to https://github.com/qubvel/segmentation_models/tree/master/segmentation_models/unet . Based on their u-net model, we combine Inception v3 from keras.application as encoder. U-net model details could be found in Model folder.https://github.com/helibu/Airbus-shipdetection/tree/master/Model/U-net%20model. We use DICE/F1 score of the intersection of predictions and ground truth based on the uneven distribution of has-ship pixels and no ship pixels. We choose U-net because most ships on images are quite small, and we need a model that can handle high resolution information. So, we choose U-net.
we have finished the framework of the web using html and css.
We used Flask to deploy our Python language machine learning model to our web application.
We use AWS elastic beanstalk service to run our web app, choosing python as the ec2 instance environment language, and also have applied for a domain name for our website. Our website address is: http://www.shipdetection.com/
This is the homepage of our website. We can upload a bunch of images in the same time, and see the image by click the "Images" button.
In our image page, we can view our uploaded images with "has ship" or "no ship" on them, and if there is a ship on the image, the ship will be bordered by a red border-box. Also there is a "Clear images" button to clear the result.
