BFEA: A SAR Ship Detection Model Based on Attention Mechanism and Multiscale Feature Fusion

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 17; pp. 11163 - 11177
• Main Authors: Zhou, Liming, Wan, Ziye, Zhao, Shuai, Han, Hongyu, Liu, Yang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Aggregation; Convolution; Datasets; Deep learning; Feature extraction; feature fusion network; Image resolution; Interference; Marine vehicles; Modules; multiscale ship detection; Object detection; Radar polarimetry; SAR (radar); Semantics; Ships; Synthetic aperture radar; synthetic aperture radar (SAR); Target detection
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2024.3408339

With the advancements in deep learning and synthetic aperture radar (SAR) technology, an increasing number of individuals are utilizing deep-learning techniques to detect ships in SAR images. However, the efficiency of SAR ship detection is affected by complex background interference and various ship sizes. Addressing these challenges, this article proposes a balanced feature enhanced attention model. First, we introduce a novel attention feature fusion network (WEF-Net) tailored for SAR multiscale ship detection. WEF-Net effectively balances the information across different backbone layers and harmonizes semantic information from various levels of the feature pyramid through aggregation and averaging. Next, we embed the receiving field extension module in WEF-Net to learn the context information and generate the global characteristics of the receiving field balance. In addition, it can extract features from multiple scales to enhance the detection capability of the model for ships of different scales. At the same time, acknowledging the impact of surrounding complex background interference on the detector, we redesigned the ELAN module by combining convolution and attention. This enhancement enables the model to better attend to target position information during feature fusion, suppress the surrounding complex background interference, and highlight the ship's feature information. Finally, owing to the prevalence of small targets in SAR images, we employ an optimized loss function to bolster the model's performance in detecting small targets. This approach accelerates training convergence, reduces instances of missed detection on small targets, and enhances overall detection performance across multiple scales. Experimental results demonstrate that our model achieves detection accuracies of 98%, 93.1%, and 76.9% on the SAR ship detection dataset, high-resolution SAR image dataset, and large-scale SAR ship detection dataset, respectively, effectively discerning ship targets amid complex backgrounds in SAR images.