Leida xuebao (Apr 2024)

Few-shot Ship Classification of SAR Images via Scattering Point Topology and Dual-branch Convolutional Neural Network

  • Yipeng ZHANG,
  • Dongdong LU,
  • Xiaolan QIU,
  • Fei LI

DOI
https://doi.org/10.12000/JR23172
Journal volume & issue
Vol. 13, no. 2
pp. 411 – 427

Abstract

Read online

With the widespread application of Synthetic Aperture Radar (SAR) images in ship detection and recognition, accurate and efficient ship classification has become an urgent issue that needs to be addressed. In few-shot learning, conventional methods often suffer from limited generalization capabilities. Herein, additional information and features are introduced to enhance the understanding and generalization capabilities of the model for targets. To address this challenge, this study proposes a few-shot ship classification method for SAR images based on scattering point topology and Dual-Branch Convolutional Neural Network (DB-CNN). First, a topology structure was constructed using scattering key points to characterize the structural and shape features of ship targets. Second, the Laplacian matrix of the topology structure was calculated to transform the topological relations between scattering points into a matrix form. Finally, the original image and Laplacian matrix were used as inputs to the DB-CNN for feature extraction. Regarding network architecture, a DB-CNN comprising two independent convolution branches was designed. These branches were tasked with processing visual and topological features, employing two cross-fusion attention modules to collaboratively merge features from both branches. This approach effectively integrates the topological relations of target scattering points into the automated learning process of the network, enhancing the generalization capabilities and enhancing the classification accuracy of the model. Experimental results demonstrated that the proposed approach obtained average accuracies of 53.80% and 73.00% in 1-shot and 5-shot tasks, respectively, on the OpenSARShip dataset. Similarly, on the FUSAR-Ship dataset, it achieved average accuracies of 54.44% and 71.36% in 1-shot and 5-shot tasks, respectively. In the case of both 1-shot and 5-shot tasks, the proposed approach outperformed the baseline by >15% in terms of accuracy, underscoring the effectiveness of incorporating scattering point topology in few-shot ship classification of SAR images.

Keywords