Ship detection by modified RetinaNet

Ship detection in optical remote sensing imagery has been a hot topic in recent years and achieved promising performance. However, there are still several problems in detecting ships with various sizes. The key objective of all scales precise positioning is to obtain a high resolution feature map while having a high semantic characteristic information. Based on this idea, a modified RetinaNet (M-RetinaNet) is proposed to build dense connections between shallow and deep feature maps, which aims at solving problems resulting from different sizes of ships. It consists of a baseline residual network and a modified multi-scale network. The modified multi-scale network includes a top-down pathway and a bottom-up pathway, both of which build on the multi-scale base network. The benefits of this model are two folds: first, it can generate feature maps with high semantic information at each layer by introducing dense lateral connections from deep to shallow; second, it maintains high spatial resolution in deep layers. Comprehensive evaluations on a ship dataset and comparison with several state-of-the-art approaches demonstrate the effectiveness of the proposed network.