|
In the existing high resolution optical remote sensing images, objects are often located in a complex environment and most datasets have serious problems of imbalance in the number and size of samples, especially the ship samples which located in a changeable marine environment. Our research is mainly to solve the problem of object detection of different scales in unbalanced datasets in complex background. Ship datasets are typical datasets with these characteristics, so we will choose ship targets as experimental dataset. We introduce a Multi-scale Graph Convolutional Network(MGCN), which is formed by a multi-scale module and GCN module to get a better performance on ship detection. For the multi-scale module, we add dilated convolutions combinations with suitable dilated rates on feature maps of different sizes to obtain context and global information, which improves the detection accuracy of multi-scale objects. For the GCN module, we try to design a co-occurrence matrix as the input of GCN to summarize the relationship from the dataset as the prior knowledge. By updating features from related objects, it can enhance local representations to get more accurate result. MGCN outperforms than existing methods on ship detection and provides a new baseline for the dataset. Experiments verify the effectiveness of our method, e.g. achieving around 16.7\% on ship detection dataset FGSD in terms of mAP. We also visualize ship detection results and show the improvement of our method. Our network is generalized and can be applied to different types of datasets. |
|
Keywords:Ship Detection; Graph Convolutional Networks; Multi-scale objects; High resolution optical remote sensing image |
|