Marine Target Detection by Exploiting Multi-Circle Features with Convolutional Neural Network

Marine target detection with coastal radars is challenging, due to the strong sea clutter and the sidelobe of land clutter and city clutter. The non-uniform non-stationary and non-Gaussian clutter degrades the performance of traditional power-based threshold detection methods based on the statistical signal processing theory. Most recent works on the feature-based target detection methods relies on the time-frequency features, which may be unavailable for short-dwell-time applications. Therefore, this work develops a new multi-dimensional feature of targets, multi-circle feature, without limitation of the dwell-time length. A convolutional neural network (CNN) is employed to discriminate true targets and false alarms after the low-threshold detection in energy domain. The experimental results exhibit that the proposed CNN-based detection can achieve a nearly 80% reduction in false-alarm rate without the loss in detection rate compared to the traditional threshold detection.