A Lightweight SAR Ship Detector Based on Whole Process Collaborative Designing

The deep learning models used in SAR ship detection always have a large model size and are computation intensive, which hinders the applications in real-time scene. In this paper, we aim to solve this problem by the whole process collaborative designing. The whole process incorporates the backbone, neck, head, label assignment and loss function. They are improved and optimized to best balance the accuracy and latency of the detector. The backbone, neck and head are improved for fusing features better according to the special characteristics of SAR images so as to increase the accuracy and reduce the latency. The anchor, label assignment and loss function are improved for utilizing samples better according to the characteristics of sparse distribution of ships so as to increase the accuracy. The proposed method achieved 97.6%, 91.2%, 91.2% and 93.6% AP50 with 108 FPS on SSDD, SAR-Ship-Dataset, AIR-SARShip and HRSID respectively. It shows great superiority in accuracy and latency compare with the other advanced detectors. This study shows that the proposed whole process collaborative designing can realize real-time SAR ship detection compare with the individual innovation. The future researchers can design the lightweight according to their demand based on the findings in this paper.