Wei Fanming, Wang Xiao
College of Computer and Information Engineering, Nanjing Tech University, Nanjing 211816, China.
Sensors (Basel). 2023 Sep 3;23(17):7641. doi: 10.3390/s23177641.
The advancements in ship detection technology using convolutional neural networks (CNNs) regarding synthetic aperture radar (SAR) images have been significant. Yet, there are still some limitations in the existing detection algorithms. First, the backbones cannot generate high-quality multiscale feature maps. Second, there is a lack of suitable attention mechanisms to suppress false alarms. Third, the current feature intensification algorithms are unable to effectively enhance the shallow feature's semantic information, which hinders the detection of small ships. Fourth, top-level feature maps have rich semantic information; however, as a result of the reduction of channels, the semantic information is weakened. These four problems lead to poor performance in SAR ship detection and recognition. To address the mentioned issues, we put forward a new approach that has the following characteristics. First, we use Convnext as the backbone to generate high-quality multiscale feature maps. Second, to suppress false alarms, the multi-pooling channel attention (MPCA) is designed to generate a corresponding weight for each channel, suppressing redundant feature maps, and further optimizing the feature maps generated by Convnext. Third, a feature intensification pyramid network (FIPN) is specifically designed to intensify the feature maps, especially the shallow feature maps. Fourth, a top-level feature intensification (TLFI) is also proposed to compensate for semantic information loss within the top-level feature maps by utilizing semantic information from different spaces. The experimental dataset employed is the SAR Ship Detection Dataset (SSDD), and the experimental findings display that our approach exhibits superiority compared to other advanced approaches. The overall Average Precision (AP) reaches up to 95.6% on the SSDD, which improves the accuracy by at least 1.7% compared to the current excellent methods.
利用卷积神经网络（CNN）进行合成孔径雷达（SAR）图像船舶检测技术取得了重大进展。然而，现有检测算法仍存在一些局限性。首先，主干网络无法生成高质量的多尺度特征图。其次，缺乏合适的注意力机制来抑制误报。第三，当前的特征增强算法无法有效增强浅层特征的语义信息，这阻碍了小目标船舶的检测。第四，顶层特征图具有丰富的语义信息；然而，由于通道数减少，语义信息被削弱。这四个问题导致SAR船舶检测与识别性能不佳。为了解决上述问题，我们提出了一种具有以下特点的新方法。首先，我们使用Convnext作为主干网络来生成高质量的多尺度特征图。其次，为了抑制误报，设计了多池化通道注意力（MPCA），为每个通道生成相应的权重，抑制冗余特征图，并进一步优化Convnext生成的特征图。第三，专门设计了一个特征增强金字塔网络（FIPN）来增强特征图，特别是浅层特征图。第四，还提出了一种顶层特征增强（TLFI）方法，通过利用来自不同空间的语义信息来弥补顶层特征图中的语义信息损失。所采用的实验数据集是SAR船舶检测数据集（SSDD），实验结果表明，我们的方法与其他先进方法相比具有优越性。在SSDD上，总体平均精度（AP）达到95.6%，与当前优秀方法相比，准确率至少提高了1.7%。
