Ship Detection Methods in Large Scene SAR Images Based on Lightweight Optimization and Knowledge Distillation
Zhijie Luo, Zeyu Yan, Taoran Jiang · 2024
Abstract: The process of identifying ships in satellite scanning images is known as ship detection methods in SAR images, which consistently scan vast areas of the ocean. Clearly, due to the fact that the images are scanning vast scenes, the size of the ships within these scenes is generally small. This poses a significant challenge for the target detection task. In this paper, we explore YOLOv5, YOLOv7, YOLOv8, and transformer based detector RT-DETR on ship detection in SAR images using the LS-SSDD-v1.0 datasets. Additionally, aiming at the problem of detection speed and model complexity, we design model lightweight optimization and knowledge distillation. Specifically, we have integrated a more lightweight network, MobileNetV3 based on the backbone of YOLOv5s. Thus, the modified network has fewer parameters and more complex than original one. Besides, we insert the SE attention module to retain accurate spatial position information. Moreover, a strategy, which is used to solve the disadvantage of inadequate feature retrieval capability in the YOLOv5s network, is designed that a small network can like large network has the incorporating the information ability through knowledge distillation. This strategy accelerates the convergence speed of training. Experimental results shows that both knowledge distillation and lightweight optimization are effective for SAR ship detection.