Coarse-to-fine ship detection using visual saliency fusion and feature encoding for optical satellite images

In order to overcome cloud clutters and varied sizes of objects in high-resolution optical satellite images, a novel coarse-to-fine ship detection framework is proposed. Initially, a modified saliency fusion algorithm is derived to reduce cloud clutters and extract ship candidates. Then, in coarse discrimination stage, candidates are described by introducing shape feature to eliminate regions which are not conform to ship characteristics. In fine discrimination stage, candidates are represented by local descriptor-based feature encoding, and then linear SVM is used for discrimination. Experiments on 60 images (including 467 objects) collected from Microsoft Virtual Earth demonstrate the effectiveness of the proposed framework. Specifically, the fusion of visual saliency achieves 17.07% higher Precision and 7.23% higher Recall compared with those of individual one. Moreover, using local descriptor in fine discrimination makes Precision and F-measure further be improved by 7.23% and 1.74%, respectively.