Self-supervised Representation Learning for the Object Detection of Marine Radar

Lingyu Si, Gang Li, Changwen Zheng, Fanjiang Xu

Abstract: One of the main challenges in object detection tasks is to detect the small and messy targets in complex backgrounds. For the marine radar images, this challenge is further aggravated. Compared with the generally visible light images, marine radar images consist of many small targets which are very similar to the sea cluster and difficult to be detected. To solve this problem, we proposed a method that focuses on detecting small ship targets in marine radar images and improving the detectors’ ability by using unlabeled data. Firstly, we propose a new dataset called marine radar datasets that contains an unlabed marine radar dataset (UMRD) and a labeled marine radar detection dataset(MRDD). Furthermore, based on these two datasets, we propose a new method called Self-Supervised Representation Learning Detection(SSRLD) for the marine object detection of radar images. SSRLD firstly learns initialized weights for the feature extraction network based on self-supervised learning. Secondly, we propose an object detection workflow towards marine radar based on this parameter initialization. The detection results on the marine radar detection dataset(MRDD) show the effectiveness of our proposed method. Specifically, SSRLD achieves a recall of 0.99 and precision of 0.95 on MRDD dataset that largely surpasses other state-of-art supervised learning-based object detectors and traditional methods.