TAMSNet: a target-adaptive multiscale detection network for fine-grained ship recognition in remote sensing images

Abstract

Fine-grained ship recognition is a core task of remote sensing imagery in the fields of maritime security and port management. However, when general detection models are transferred to this task, they face two major challenges: first, the fine-grained features of ships from a bird's-eye view in remote sensing are easily lost during downsampling, making it difficult to accurately distinguish between similar ship categories; second, the complex backgrounds of ports and densely arranged ships create strong interference, and slender ships are highly sensitive to rotation angles, which can easily cause localization errors. To improve fine-grained detection performance of rotated ships in remote sensing images, we propose an object-adaptive multi-scale detection framework. Specifically, we first designed an object-adaptive multi-scale feature pyramid network (TAMFPN), which implements cross-layer fine-grained feature compensation through an object-adaptive shallow-position enhancement module (TASPE), strengthening key discriminative information such as decks and superstructures. We also designed a multi-scale context interaction module (MCIM), which integrates multi-dilation spatial modeling and channel recalibration to effectively suppress background noise and interference from adjacent objects. In addition, to address the high sensitivity of ships to rotation angles, we propose a joint loss strategy combining Smooth L1 and KFIoU, which jointly regresses ship shape, size, and rotation angle, resolving the localization bias problem of slender ships. Experimental results on the FGSD2021 remote sensing fine-grained ship detection dataset show that this method effectively improves recognition and localization performance of fine-grained ships and targets in arbitrary orientations, while maintaining high overall detection accuracy.