Satellite Image Detection and Localization of Ships with a Convolutional Neural Network

Author Information

Clayton BirchenoughFollow

Is this project an undergraduate, graduate, or faculty project?

Undergraduate

Project Type

individual

Authors' Class Standing

Clayton Birchenough, Senior

Lead Presenter's Name

Clayton Birchenough

Faculty Mentor Name

Mihhail Berezovski

Abstract

For this study, a convolutional neural network was built to provide automatic ship detection and localization from satellite images. The challenge of solving for the parameters of the neural network is a non-convex optimization problem with acceptable solutions and many deceptively acceptable solutions. To avoid deceptively acceptable solutions, different learning rate parameters, data augmentation methods, and neuron dropout rates were explored when training the network. Additionally, the effects of the number and order of convolutional, max pooling, and fully connected layers were varied to investigate the impacts on training and results. Furthermore, for training the network, methods needed to be employed to handle the relatively large number of training images which would not fit on available RAM.

Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?

No

Share

COinS
 

Satellite Image Detection and Localization of Ships with a Convolutional Neural Network

For this study, a convolutional neural network was built to provide automatic ship detection and localization from satellite images. The challenge of solving for the parameters of the neural network is a non-convex optimization problem with acceptable solutions and many deceptively acceptable solutions. To avoid deceptively acceptable solutions, different learning rate parameters, data augmentation methods, and neuron dropout rates were explored when training the network. Additionally, the effects of the number and order of convolutional, max pooling, and fully connected layers were varied to investigate the impacts on training and results. Furthermore, for training the network, methods needed to be employed to handle the relatively large number of training images which would not fit on available RAM.