Comparative Ship Classification in Heterogeneous Dataset with Pre-trained Models

---

Résumé

In this work, we have proposed a comparative study on heterogeneous ship images classification via pre-trained models (VGG16, Inception-V3, ResNet-50). We have introduced a new dataset in this paper: heterogeneous dataset, where data were collected from two different sensors i.e. optical sensor and radar sensor. We have proposed three classes classification solutions. Our objective was to separate the ship images from the others. We selected a convolutional neural network (CNN) as the backbone network. We also used transfer learning to reduce the computational time. Among the three trained models, we have realized that ResNet-50 has the best performance with less misclassified classes and higher testing metrics.

Mots-clés

SAR images , optical satellite images , pre-trained model , ship classification