SVSDet: A Fine-Grained Recognition Method for Ship Target Using Satellite Video

Abstract

Target recognition from remote sensing images is commonly challenging because of large-scale variations and small objects, and these challenges are more prominent in satellite video images. The current object detection algorithms have some difficulties in fine-grained feature extraction and classification for multiscale and small objects. We propose a novel model called the SVSDet method based on YOLOv5 improvement to address the above-mentioned issues. In this method, we have introduced the space-to-depth module into the backbone of the network, which enhances the network's ability to extract fine-grained features. The neck structure is improved by using the bidirectional feature pyramid network to enhance the network's ability to extract features at multiple scales, thereby improving its overall multiscale feature extraction ability. Subsequently, we have replaced the C3 module in the original network's neck with the C2f module to obtain more abundant gradient flow information. This helps to improve the network's performance further. Finally, the coordinate attention module is introduced into the cross-scale feature connection path, which effectively enhances the network's target detection and recognition performance. We have conducted extensive comparative experiments and ablation experiments on the publicly available datasets ShipRSImageNet and SAT-MTB to confirm the effectiveness of our proposed SVSDet method. The performance of this approach is then evaluated using Jilin 1 satellite video data, and it outperforms the main YOLO series algorithms currently used.

Keywords

• Attention mechanism
• deep learning
• multiscale feature fusion
• satellite video
• ship recognition