Multi-scale Graph Aggregation based Ship Target Recognition

The deep learning method for ship recognition on remote sensing images has the advantages of high efficiency and high precision. However, traditional ship detection methods have two shortcomings: 1) The convolution receptive field used for feature extraction is fixed, which is difficult to be compatible with the feature extraction of multi-scale objects, resulting in the distortion of some object features. 2) Traditional pixel-based methods lack object-level integrity in identifying object contours in complex scenes. This paper proposes a multi-scale ship target recognition model (mGAT), which realizes the dynamic extraction of multi-scale object features through a multi-scale feature fusion convolution network based on gating mechanism, and combines the idea of graph aggregation in GAT to propagate the object feature information. Experiments show that, compared with traditional semantic segmentation methods, the proposed method can effectively improve the recognition accuracy of multi-scale ships and other objects in the scene.