Machine Learning for Ship Detection with Radar

Surface search radars have been widely used in the detection and location of ships, with the aim of promoting navigation safety and national defense. Although they are proven systems, one of the main challenges for these sensors is to maximize the detection probability and reduce the false alarm probability in the face of sea clutter. Taking this into account, in this work, two approaches based on machine learning are presented to improve the performance of surface search radars in scenarios with sea clutter. Both methods involve the use of Convolutional Autoencoders (CAE) to extract the spatial features of the data. The CAEs are used as filters and anomaly detectors, where the presence of ships is interpreted as an anomaly. In the first approach, a thresholding method is used to detect ships, while in the second the You Only Look Once (YOLO) object detection algorithm is applied to the images denoised by the CAE. To train, test and validate the proposed solutions, a dataset containing real radar images was built. The results show that the use of the autoencoder is able to reduce the sea clutter from the obtained radar images. Furthermore, the results indicate that the detection through YOLO outperforms thresholding detection in terms of the mean squared error (MSE) of the processed images.