现有合成孔径雷达(synthetic aperture radar，SAR)舰船实例分割方法未实现掩模交互或交互性能有限，导致检测精度较低。针对上述问题，提出了一种基于掩模注意型交互(mask attention interaction，MAI)的SAR舰船实例分割方法MAI-Net。首先，MAI-Net使用了膨胀空间金字塔池化，来获取多分辨率特征响应，增强了对背景的鉴别能力。其次，MAI-Net使用了非局部注意力模块来抑制低价值信息，实现了空间特征自注意。最后，MAI-Net提出了拼接混洗注意力模块来平衡不同特征图的贡献，进一步提高了实例分割精度。在公开的像素级多边形分割 SAR 舰船检测数据集(polygon segmentation SAR ship detection dataset，PSeg-SSDD)上的实验结果表明，MAI-Net的SAR舰船实例分割精度高于现有其他11种对比模型，实例分割精度达到61。1％，高于次优模型1。5％。
Mask attention interaction for SAR ship instance segmentation
The current synthetic aperture radar(SAR)ship instance segmentation models fail to realize mask interaction or the interaction performance is limited,resulting in low detection accuracy.To solve this problem,a SAR ship instance segmentation method based on mask attention interaction(MAI)is proposed,called MAI-Net.Firstly,MAI-Net uses the atrous spatial pyramid pooling(ASPP)to obtain multi-resolution feature responses and enhance the background identification capability.Secondly,MAI-Net uses a non-local block(NLB)to suppress useless information and realize spatial feature self-attention.Finally,MAI-Net proposes the concatenation shuffle attention block(CSAB),which can balance the contribution of different features and further improve the instance segmentation accuracy.The results on the public polygon segmentation SAR ship detection dataset(PSeg-SSDD)show that the SAR ship instance segmentation accuracy of MAI-Net is higher than that of the other eleven comparison models,the accuracy is 61.1％,1.5％higher than the suboptimal model.
万方数据
维普
