A Lightweight Ship Identification Method for Embedded Devices in Marine Radar Surveillance Systems

In scenarios such as near-shore waters and inland rivers, radar systems represent ships as a pile of moving blips that can easily be mistaken for noise, reefs and coastlines. Meanwhile, due to the performance limitations of computing equipment, it is unable to operate the ship identification models with high computing consumption smoothly. To address these issues, a novel lightweight method, namely LRNet, based on deep convolutional networks is proposed, which uses the EfficientNet_B0 model to extract key information about the ship blips in the radar images. In addition, to ensure the scale invariance of convolutional features, a Spatial Pyramid Pooling-based model is used. In the last procedure of prediction, a Cluster-NMS method and a Complete-IoU function are used to increase the identification accuracy and convergence speed. To validate the proposed method, a radar image dataset was established, while several experiments were conducted, which showed that the recall, identification accuracy and precision of LRNet reached 0.9025, 0.8837 and 0.8942 respectively. On the other hand, the parameter size and calculational consumption were 16.75M and 10.85G respectively. Compared with other standard identification algorithms, the LRNet has a superior performance in identifying ships in various scenarios under marine radar images, especially which can be effectively deployed in embedded devices.