SAR ship localization method with denoising and feature refinement

Zha, Cheng; Min, Weidong; Han, Qing; Li, Wei; Xiong, Xin; Wang, Qi; Zhu, Meng

doi:10.1016/j.engappai.2023.106444

Cited by 5 publications

(1 citation statement)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Liu et al [20] introduced coordinate attention to the YOLOv7-tiny model and improved the spatial pyramid pool (SPP) and SIoU loss function to strengthen detection performance. Cheng et al [21] used non-local mean as a denoising method for SAR images to preprocess detected images, and then proposed a feature thinning module to suppress background interference and improve the positioning performance of the model. Zhang et al [22] designed five modules to form a high-precision detection network called HyperLi-Net.…”

Section: Introductionmentioning

confidence: 99%

CSEF-Net: Cross-Scale SAR Ship Detection Network Based on Efficient Receptive Field and Enhanced Hierarchical Fusion

Zhang,

2024

Remote Sensing

View full text Add to dashboard Cite

Ship detection using synthetic aperture radar (SAR) images is widely applied to marine monitoring, ship identification, and other intelligent maritime applications. It also improves shipping efficiency, reduces marine traffic accidents, and promotes marine resource development. Land reflection and sea clutter introduce noise into SAR imaging, making the ship features in the image less prominent, which makes the detection of multi-scale ship targets more difficult. Therefore, a cross-scale ship detection network for SAR images based on efficient receptive field and enhanced hierarchical fusion is proposed. In order to retain more information and lighten the weight of the network, an efficient receptive field feature extraction backbone network (ERFBNet) is designed, and the multi-channel coordinate attention mechanism (MCCA) is embedded to highlight the ship features. Then, an enhanced hierarchical feature fusion network (EHFNet) is proposed to better characterize the features by fusing information from lower and higher layers. Finally, the feature map is input into the detection head with improved bounding box loss function. Using SSDD and HRSID as experimental datasets, average accuracies of 97.3% and 90.6% were obtained, respectively, and the network performed well in most scenarios.

show abstract

Section: Introductionmentioning

confidence: 99%