MLBR-YOLOX: An Efficient SAR Ship Detection Network With Multilevel Background Removing Modules

Abstract

On the remote sensing images of marine synthetic aperture radar (SAR), ship targets often occupy only a small part of an image, and the rest are all sea and coastal backgrounds. Existing neural networks based on SAR ship detection often directly detect an entire SAR image, which consumes a large amount of computing resources. In this article, a new marine SAR ship detection network, called multilevel background removing–you only look once X (MLBR-YOLOX), is proposed. First, a new plug-and-play module, called standalone spatial patch detector, is proposed to predetect the position of ships and filter out most of the sea backgrounds at the source image level. Second, a deep spatial feature detector is presented for detecting deep semantic features of the output of the backbone module to further reduce the computational cost of the neck and head modules. Finally, the original YOLOX network is adopted to locate and classify the pre-detected results. Experimental results on the SAR ship detection dataset and the high-resolution SAR images dataset indicate that the detection performance of MLBR-YOLOX is close to that of YOLOX, but the computational complexity is merely 23.97% and 12.50% of YOLOX's, respectively. Moreover, the experiment conducted on a large-scene Sentinel-1 SAR image illustrates that the proposed network has good migration application capability.

Keywords

• Convolutional neural network (CNN)
• multilevel background removing modules
• ship detection
• synthetic aperture radar (SAR) image