Lightweight attention-based SAR ship detector

Synthetic aperture radar (SAR) remote sensing images have been widely applied in military reconnaissance and traffic supervision, owing to their all-weather and all-day abilities. With excellent learning performance, convolutional neural networks are employed in the SAR ship detection algorithms. However, it is difficult to extract features. In practical applications, computing resources and memory space are limited, and high inference speed is required. Therefore, a lightweight attention-based ship detector (LASD) was proposed. A novel linear hybrid attention module was designed which extracted potential ship features from deep-level space by using global channel attention and local spatial attention. A spatial pyramid pooling module based on cross-stage partial connections optimized the quality of multi-scale feature fusion, which replaced the parallel max-pooling group with large kernels with the serial max-poolings with small kernels to improve the inference speed. A novel feature fusion scheme via the local channel attention was suggested which widened the gap between the objects and background noise using local attention during the feature fusion. The results on the public datasets SSDD and LS-SSDD-v1.0 show that LASD achieves the balance of detection precision and inference speed, and is more competitive than the other advanced algorithms.