Abstract
Recently, in the field of Synthetic Aperture Radar (SAR) ship detection, deep learning-based methods have made significant strides in terms of detection accuracy and speed. However, small-scale targets and complex backgrounds remain a formidable obstacle to SAR ship detection. To overcome the aforementioned challenges, this letter proposes LiteSAR-Net, a lightweight SAR ship detector, with the aim of enhancing ship detection capabilities in SAR imagery. In detail, an End-to-end Image Preprocessing Network (E2IPNet) is proposed to strengthen context information and expand the network’s effective receptive field. In addition, to prevent the dilution of semantic information, the Channel Feature Guided Spatial Pyramid Pooling (CFGSPP) is proposed, which can adjust the parameters adaptively based on inter-channel information. The proposed LiteSAR-Net achieved an Average Precision (AP) of 98.61% on the SAR Ship Detection Dataset (SSDD) and 93.33% on the High-resolution SAR Images Dataset (HRSID), with a parameter of only 5.247M, outperformed many state-of-the-art (SOTA) detectors. Our code can be found at https://github.com/ZYMCCX/LiteSAR-Net.
Funding Information
  • Natural Science Foundation of Jiangsu Province (BK20211365)
  • National Natural Science Foundation of China (62076107)

This publication has 15 references indexed in Scilit: