Multiscale Accurate Ship Detection Network Driven by Multiattention Fusion for Complex Maritime Backgrounds

Ship detection based on solid-state pulse compression radar (SSPR) plays a key role in the development of smart shipping, which can effectively ensure safe and efficient navigation. However, SSPR image suffers from inherent limitations such as inconspicuous ship features, large differences in ship scales, and complex backgrounds; and most of the existing methods are anchor-based, which is susceptible to anchor overlap and feature suppression. To address the above limitations, this article customizes a complementary scheme of multiscale encoding and multiattention fusion, called MMAF-Net, aiming to enhance the accurate detection of multiscale targets in complex maritime background environments. Specifically, CenterNet is first employed as the encoding backbone to model the target as a single centroid to avoid anchor overlap and feature suppression. Second, a pixel shuffling-assisted feature pyramid module is embedded to enhance the comprehensive perception of multiscale ship features. Finally, a multiattention fusion module (MAFM) is designed, which can selectively focus on small-size ship features while suppressing other background interfering features. Thanks to the organic integration of the above modules, the comparative validation on the recognized SSDD and the self-constructed SSPR-ship dataset shows that the proposed method is highly competitive in terms of accuracy and real-time performance.