基于显著性CNN的SAR图像靠岸舰船检测方法

A saliency-based CNN method for inshore ship detection in SAR images

为提高卷积神经网络(CNN)对SAR图像靠岸舰船的检测精度,提出一种基于显著性CNN方法.该方法使用视觉显著性机制对SAR图像进行预处理,将得到的场景注意力加权(即显著图)融合到原始SAR图像中,最终将带有场景注意力加权的SAR图像输入到CNN网络.在公开的SAR舰船检测数据集上的实验表明,与经典双阶段检测器FasterR-CNN方法相比,显著性CNN方法可抑制岸边背景干扰,有效提高SAR靠岸舰船的检测精度.

In order to improve the accuracy of convolutional neural network(CNN)for detecting inshore ships in SAR images,this paper proposes a saliency-based CNN method.In this method,the visual salience mechanism is used to preprocess SAR images,then the obtained scene attention weight(i.e.,salience graph)is fused into the original SAR images,and finally the SAR images with scene attention weight is input into the CNN network.Experiments on public SAR ship detection datasets show that,compared with the classical two-stage detector Faster R-CNN method,the saliency CNN method can suppress the shore background interference,and effectively improve the detection accuracy of SAR docked ships.