A Lightweight SAR Ship Detection Network Based on Deep Multiscale Grouped Convolution, Network Pruning, and Knowledge Distillation

Abstract

Deep learning has proven to be highly effective in synthetic aperture radar (SAR) image target detection. However, many latest deep learning models have predominantly focused on increasing depth and size to enhance detection accuracy, often ignoring the balance between accuracy and detection speed, as well as the practical deployment of these models on hardware platforms. Therefore, a lightweight algorithm for SAR ship detection is designed in this article. First, we propose a preliminary lightweight scheme, including a multiscale feature learning augmented backbone, a lightweight feature fusion neck, and a parameter-sharing lightweight detection head. Second, unimportant branches of the network are pruned to further compress the model. Finally, the detection accuracy of the model is enhanced by knowledge distillation without augmenting the model volume, which compensates for the accuracy loss caused by model compression. Experimental validation is conducted on three SAR image ship detection datasets (SSDD, high-resolution SAR images dataset, large-scale SAR ship detection dataset-v1.0) to thoroughly assess the effectiveness of the proposed lightweight algorithm. Experimental results on the three datasets demonstrate that the proposed method achieves a model volume reduction to one-third of the baseline while maintaining a minimal decrease in detection accuracy. In SSDD, the proposed method achieved 98.7 accuracy, 0.92M parameters, 3.1G FLOPS and 2.1 MB size of 1.5X pruning rate. Furthermore, it outperforms other state-of-the-art lightweight detectors.

Keywords

• Convolutional neural network (CNN)
• knowledge distillation (KD)
• lightweight
• network pruning
• synthetic aperture radar (SAR) target detection