Multigranularity Self-Attention Network for Fine-Grained Ship Detection in Remote Sensing Images

Abstract

Fine-grained ship detection is very important in the remote sensing field. Most previous remote sensing object detection works only utilize the global features for fine-grained object detection, which ignores the local information, deteriorating the detection performance. In this article, we propose a multigranularity self-attention network (MGANet), which can exploit both the global and local features for fine-grained ship detection. The MGANet consists of two modules: a local-global features alignment module (LAM) and a multigranularity self-attention module (MSM). The LAM is designed to align features of object parts and features of the object by using the convolution with different strides. The MSM introduces a self-attention mechanism that can effectively fuse the global and local features for fine-grained ship detection. In addition, we launched an oriented bounding box-based fine-grained ship detection (OFSD) dataset which is the largest fine-grained ship dataset to test and verify the effectiveness of the proposed MGANet method. Comprehensive evaluations on the OFSD and ShipRSImageNet datasets demonstrate the superiority of our proposed MGANet method over existing state-of-the-art methods for fine-grained ship detection in remote sensing images.

Keywords

• Deep learning
• fine-grained ship dataset
• fine-grained ship detection
• multigranularity
• self-attention mechanism