A proposal learning strategy on CNN architectures for targets classification

Synthetic Aperture Radar (SAR) imagery has a great potential in remote sensing field in order to maintain aerial, land and maritime surveillance. To improve this potential for several goals, Machine Learning (ML) techniques are used for effective and efficient classification of remotely sensed imagery. The strengths of machine learning include the capacity to handle data of high dimensionality with several modality and to map classes with very complex and varied characteristics.We focus in this paper on the ship classification from SAR imagery using a deep learning algorithms with transfer leaning mechanisms. To improve the recognition performances, several learning strategies are proposed. We illustrate these issues through applying a transfer learning on deep learning ship classification with several learning strategies. We used in this work especially no huge labelled dataset available for training. The aim is thus to see how to leverage knowledge from models pre-trained on other tasks (source tasks) and use them for target classification (target task).