Projects

Airbus & Ship Detection

The product is to utilize satellite images provided by Airbus and machine learning model to detect ships from satellite images with a fast speed.

Data analysis

With data analysis, we found out that most dataset have no ships. And pixels of those which have ships are pretty small. From this, we thought it’s better to have a model to tell the ship existence and another model can process segmentation information with high resolution.

Machine Learning Model and Results

As said in the data analysis part, two models were developed.

One is to tell if there is ship on the image, which is a ship-existence model. This model is trained with whole dataset. An ensemble model combining pretrained ResNet50 and VGG16 with weights of ImageNet is used here.

We used a four-layer CNN model at first, and then we tried pretrained ResNet50, ResNet34, DesNet169, Inception v3, Inception ResNet v2 and VGG16. We found out that an ensemble model, as in Fig 3, combining ResNet50 and VGG16 gets the best result as Fig 4. Both the training and validation accuracy are around 95%.

Google Sites

Report abuse