CNN Models Transfer Learning Based Ship Classification for Intelligent Ship Routing

Collision free autonomous navigation is challenging area of interest for the shipping industry, and an accurately classifying the ships according to their size and type classes is a crucial undertaking in marine operations. Accurate prediction and identification will improve the maritime security and navigation safety. However, automatic ship classification is a difficult subject because of the wide variety of ship designs and the complexity of visual data. We developed a deep learning model using transfer learning from pre-trained models in order to classify five different kinds of ships. We performed the experiment using different pre-trained convolution neural network model of the deep learning. In the transfer learning we considered the training of the dataset for 100 epochs. In this study, we investigated four deep learning model ResNet 50, VGG 16, VGG 19, and Inception V3. We carried out the identical procedure for training of each model by removing the top layer of pre-trained model and adding three dense layers for the top layers. The neuron counts are 1024, 512, 256 respectively followed by dropout layer of 0.5. Results of our model shows the maximum obtained accuracies are 86.66 % and from these we can observed that the developed model can be used for ship identification and classification and to take the decisions for collision free and smooth navigation.