LRSNet: Lightweight and Real-time SAR Ship Detection with Semantic Segmentation
Dong Chen, Yanwei Ju · 2021 CIE International Conference on Radar (Radar) · 2021
Object detection based on deep learning methods plays a significant role in usual tasks. These greatly universal detection algorithms can promote the intelligent interpretation of synthetic aperture radar (SAR) images and help to get more accurate detection results. Different from the current methods directly used in SAR images, we creatively try to apply the semantic segmentation idea to SAR ship detection and complete detection and segmentation at the same time. We think SAR ship detection as a pixel-by-pixel binary classification problem rather than common general detection methods. By this way, the problem of multi-scale ships and small targets can be transformed into the unbalance of foreground and background in SAR images. In this paper, we propose a lightweight but efficient convolutional neural network (CNN) model named LRSNet. LRSNet has fewer parameters, higher efficiency and better detection and segmentation results. On HRSID dataset, the model can achieve 42.6% AP and 59.5% AR, which is better than the yolov4 detection algorithm, and the model size is only 8.4M.