Network Slimming Method for SAR Ship Detection Based on Knowlegde Distillation

This paper proposes a network slimming method for synthetic aperture radar (SAR) ship detection based on knowledge distillation. Firstly, the generic objection detection network is pruned regularly and extremely on filter-level to get lightweight models under different global pruning ratios. Secondly, the knowledge distillation framework based on Kullback Leibler (KL) Divergence is used to train small student network and large teacher network from scratch synchronously to restore the accuracy of student network. To verify the effectiveness of the proposed method, sufficient experiments are conducted on the widely used SAR Ship Detection Dataset (SSDD). YOLO [email protected] is selected as the baseline model while YOLO [email protected] as the teacher network. Results show that, with our method, a student model with only 15.4M parameters (25% of the baseline model) can achieve high pruning ratio while still maintaining encouraging performance. Compared with the baseline model, there are only 1% and 0.9% differences on average precision (AP) and average recall (AR), respectively. Compared with traditional fine-tuning method which only restores 0.6% AP, the model slimming method based on knowledge distillation proposed in this paper restores 2.4% AP with obvious advantages.